Repository navigation
Status Update Log
Scope: Dated update notes moved verbatim from Project status and Runtime validation status (2026-11-06) so those pages stay current-state summaries. Each note keeps its original wording and evidence label; current maturity and evidence scope remain owned by the two status pages and the canonical domain owners.
2026-10-27 owner-requested Wind/XP/debug changes — Runtime-confirmed bounded; promotion approved. Correct WIND2/WIND3 portrait order, remove stage-specific XP limits in both progression modes, and add an unchecked Debug options → Start with all stage keys setting at the bottom of the browser settings. All 21 keys/crystals are added alongside normal supplies and restored on fresh-run reset within existing inventory/lifecycle allocations. The maintainer accepted proof v01 (SHA-256 7129c455a29f736c77bb06b6a34d5695326f6eae48981b2a82a8739e47bc8465) and explicitly approved PR #174 merge. This is bounded user-reported validation, not exhaustive stage/seed coverage; audio investigation status is unchanged. Evidence and reproduction are in docs/key-icons-xp-debug-validation.md.
Focused selector RSP/task-event upstream static audit (2026-10-11; no emulator run): In five complete v05 captures, at VI #2165 the failing run's second RSP execution/completion is observed at VI+7,376 Count and SP_INT at +7,396, before AI_LEN at +8,266; three heavy-input non-reproductions complete at +7,720/+7,990/+7,950 and interrupt at +7,732/+7,998/+7,966, after AI_LEN at +7,594. All first executions are VI+6,976. Pinned RMG do_SP_Task and ParaLLEl RSP source statically establish that the first execution returns running, schedules a zero-modeled-delay RSP_TSK_EVT, and the second execution returns HALT|BROKE|INTR_BREAK|TASKDONE. Thus the earlier local divergent observation is delayed continuation-event dispatch, not proven differing RSP microcode cost. Pre-read SP is neither necessary nor sufficient for the later FIFO state: 193/299 failing FULL services have pre-read SP; the no-repro minimal-input recording has 3,368 such services and zero FULL, while isolated failures in three healthy controls have no pre-read SP. Other feedback/phase differences existed from VI #294/#295 and an SP ordering split at VI #601 caused no PCM size change, so the 1,872nd PCM attempt is the first common size fork, not the first system divergence. Static scheduler points at AudioInfo OSTask type 2 with active pointer 0x8013BE28 but RSP task type and guest PC/EPC were not recorded. Guarded selector/audio fixed ROM spans are disjoint; whole-composition memory/ABI and unknown wild writes remain Pending. Release audio gate BLOCKED. Full report archived in Research canonical investigation.
NEW earliest selector audio-size divergence (2026-10-10; five complete v05 captures, source/trace-confirmed, origin Pending): Aligning native PCM attempts across the earlier 299-rejection frontend failure and three heavy-input no-runaway captures proves identical first 1,872 submitted PCM sizes; at attempt index 1,872, failing run selects 1,472B while all three choose 1,408B. Exact preceding AI_LEN feedback 1,344B vs 1,352B; the native stock size formula maps those to 368 vs 352 frames. Pinned RMG get_remaining_dma_length() quantizes an unrounded 1,351.728B vs 1,353.053–1,353.789B down to the 8B boundary. Failing AI_LEN service is 8,266 Count after VI, three controls 7,594 Count after VI; similar service-offset excursions also occur in healthy traces, so this is a phase-sensitive threshold, not proof of a specific sound/helper operation. The minimal-input case diverges earlier at attempt 1,848 and is not part of this matched prefix. The RMG fractional-duration diagnostic #161 does not guarantee recovery: 1,472B tail duration would increase 811,072→812,306 Count on the same assumed schedule. Source input is current ROM's native MIPS, pinned RMG C, and our Python patch pipeline; no original MKMSZ C source exists. No code correction approved, no reason to waive the release blocker or infer absence of memory corruption. Report and hashes in Library selector negative control folder.
NEW v05 selector negative controls (2026-10-10; four user-reported non-reproductions, trace-confirmed): Four complete six-stream RMG v05 archives (24/24 rings saved=total, no overwrites) contain 1,133 Up/Down PIF rising edges yet only 0, 1, 1, and 1 isolated native PCM rejects, versus 299 recurrent rejects in the earlier failing menu run with 245 vertical edges. Minimal-input run has zero; 239-Down, 335-Down, and 558-vertical-press runs each have exactly one, then recover. All four have zero SDL drops/failed pushes. At each isolated reject FIFO head/tail is 1,536/1,408B and queued 1,408B DMA finishes 30,336–31,064 Count BEFORE next native audio service, whose AI_STATUS is EMPTY; prior failing head/tail 1,600/1,472B, with tail 4,746 Count AFTER next service and recurrence. Even AI_LEN=1216/accepted 1,600B precedes the three healthy recoveries. This directly disproves input count or one FULL as sufficient for sustained acceleration; the original source of the phase/PCM size choice remains Pending, not resolved. All four preserve one active-VI native service per frame and exact synth sample-clock steps. No ROM/emulator changes or new tests; 1.0 audio gate still BLOCKED. New raw ZIPs and findings report archived in Library MKMSZR Research/07 - RAM Diffs and Experiment Data/RMG-v05-Stage-Select-Four-Negative-Controls-2026-10-10/. Canonical analysis.
Selector-only RMG v05 FIFO failure (2026-10-10) — Trace-confirmed; first frontend initiator Pending. Six new recordings recovered from the existing archived ZIP all match SHA manifests and have saved=total, overwritten=0: VI 13,500, RSP 20,196, phase 3,085, input 1,610, SDL 11,720, AI 17,588. During rapid Safe Stage Select navigation, 299 unique native PCM submissions are rejected t36.496–56.364; 169 precede the first stage-9 sample (t47.696, following Start t47.615), 130 follow. Zero SDL threshold drops or failed pushes (5,860 successful); this is the same native FIFO rejection class seen in previous accelerated Inventory captures, starting before Fortress or rich Inventory. First rejection: active/queued 1,600/1,472B and promoted 1,472B tail due 4,746 CP0 Count after next native service (1,408B would be done 30,518 Count beforehand). 3,081 native audio services map 1:1 to active VIs; 3,081 native phase clock steps exactly match AudioInfo frame sizes; synthesis exceeds 368/frame by 105,248 = 299 × 352 samples during the burst. PIF has 95 Up / 150 Down press edges (201 before first rejection), but does not record native input-consumption PC or SFX voice creation. Prior helper-OFF Trace 9 (blue retained) was healthy; sound-JAL-OFF Trace 10 and sound-off/s0-to-t9 Trace 11 still failed. The exact blue edit is a 29-entry BGR555 palette-only copy, separate from the helper hook. No upstream instruction causality, ROM/emulator change or production correction; 1.0 audio gate BLOCKED. See the canonical audio investigation and new report/audit archived in Library MKMSZR Research/07 - RAM Diffs and Experiment Data/RMG-v05-Stage-Select-Prestage-2026-10-10/. Earlier A–E Fortress states remain separate; E was healthy.
Long v05 negative control (2026-10-10): high Inventory/combat/input activity is NOT sufficient for sustained accelerated music; speculative off-screen Grunt AI is not instrumented. Maintainer reports a ~10-minute stress run (walking, un-killed enemy fighting, heal in another box, intensive cursor/box switching), intermittent crunch but no sustained speedup. Input/AI/RSP channels are complete (9,766 PIF, 213,789 AI, 191,216 RSP), while VI, phase and SDL rings lose 12,668 / 2,877 / 11,422 earliest records (do not call this a lossless six-stream capture). 18 actual native PCM rejects, worst 5 in 0.266s, and at least 59 retained host queue-threshold drops, vs prior v05 fast 86 in 5.734s. 348 Down + 264 Up press edges, no runaway; 10 guest rejects +23 host drops after last input, with a 167.38s entirely input-neutral ending. Three healthy 368-only plateaus (44.785, 40.000 and 38.484s), the last without a single button press, all walk AI_LEN from 1,344 to 1,288; 29 occasions of AI_LEN=1216 and 30 accepted 1,600B submissions also do not force runaway. Retained 32,768/32,768 VI callbacks each have an audio first-status service, native phase sample-clock transitions 32,767/32,767 exact. Offscreen Fortress GRUNT1 0x0E / GRUNT2 0x0F activity could perturb scheduling but none of the six streams records aggro/AI/position/RNG, so enemy blame remains Hypothesis. No production patch or further unfocused user test. Canonical detail.\n\nRMG v05 guest-PIF input → audio first-failure correlation (2026-10-10): measured but original cause Pending. New maintainer Runtime-observed clipping/speedup after Inventory Up/Down navigation. Six host traces complete/no loss; 1,594 valid port-0 PIF responses, 3,962 complete audio services with no active-VI gaps. Fortress stage 9 at t13.127s; Z response t38.786s (probable Inventory open, not an internal-state marker); 39 Down + 34 Up presses from t39.711s preceded the first native FULL at t62.544s. The first AI_LEN 1,216 t62.093997 preceded a contemporaneous Up press by 0.535ms, causing first 1,600B generation at t62.111; feedback low band stepped from 1,256→1,248→1,240→1,232→1,224→1,216 in ~4.804s / 8B (consistent with ~15.29 Count/VI drift). First full FIFO active/queued 1,600/1,472B; promoted 1,472B tail stays active 4,056 Count beyond next native service, nearly duplicating previous v04 fast 4,126 margin. 86 rejects over 5.734s among 345 services (259 accepted), producer sample clock +23.54% vs nominal frame count; zero SDL threshold drops. Last controller release t63.395; FIFO rejection continues 4.883 seconds without inputs. Thus continued movement is not necessary to sustain the fault, though Inventory opening/navigation could affect initial phase. No game instruction origin or approved correction yet; no further unspecific tests. Detailed owner; raw ZIP-independent six streams + audit preserved in Library.\n\nPending controller-input attribution (2026-10-10): v04 traces cannot tell exactly when Inventory cursor Up/Down input was received. Maintainer reports acceleration/clipping ~5–10s after renewed Inventory navigation in latest speedup, but no input stream existed to verify it. New read-only pinned RMG PIF guest-visible controller response capture (post-process/netplay, per-poll raw buttons/stick and VI-aligned host timestamp) is staged in draft diagnostic PR #165, based on v04 #164; build and maintainer parity/one bounded cursor sweep remain Pending. Do not infer game cursor-consumption instruction or physical SDL timestamp from PIF responses. Existing clean/slow/fast raw captures cannot be backfilled. No ROM patches or production change. Focused owner.\n\nv04 DMA deadline closure refinement (2026-10-10): 368 native sample-frames = 1,472B = exactly 811,072 RMG Count, just 20 Count shy of VI's 811,092; adjacent 64B output changes completion deadline by 35,264 Count. The clean and slowdown captures each spend ~40s in healthy 368-only output; DMA END cadence averages 811,077.40/811,077.75 Count against native service cadence ~811,092, producing natural ~14–15 Count/VI phase creep and a 1,344→1,280-byte AI_LEN staircase. At the slow fork, next audio service occurs 35,580 Count after 1,408-byte queued DMA completion, establishing a fresh 352-frame BUSY→EMPTY rhythm; guest FIFO is empty for median 35,226 Count between completion and next service. At fast fork a 1,472-byte queued DMA survives 4,126 Count beyond next service; if it were 1,408B it would have completed 31,138 Count before. The severe 81-rejection episode includes successful submissions between failures; it is not 81 consecutive failed services. This fully explains opposite feedback given observed timings, not which original action caused the phase choice. Read-only diagnostic in draft PR #162; 1.0 audio gate remains blocked. Detailed owner.\n\nThree new RMG v04 captures (2026-10-10) reproduce both SLOW and FAST music anomalies with complementary native audio feedback fixed points; upstream cause still Pending, release audio gate BLOCKED. Clean 379.61s: five isolated/3-cluster FIFO rejects, music normal; also a 39.77s all-368-frame plateau, proving the historical 368-only phase is not sufficient for acceleration. Slow 229.80s: only five isolated guest rejects, but at t191.400–229.801 2,303 consecutive accepted 352-frame buffers (1,408 bytes) after an intervening rejection; 2,304 accepted in wider interval, AI_LEN≈1400, −4.35% synthesized time per VI, host SDL queue median only 504B (no host drops during plateau). A DMA END only 10 CP0 Count after FULL-status read at t191.417 initiates/maintains the low-supply feedback. Fast 277.60s: 81 repeated FIFO rejections (~one per four audio services) t267.047–272.381, synthesis +23.8% versus per-VI target during burst while skipping PCM submissions, near-full host queue median 16,600B and four host drops. Across 52,316 new producer snapshots, configured clocks/rates, manager gates, AudioInfo/command addresses and synth pointer are invariant; all 52,313 sample-clock increments match the native buffer counts; stable Fortress HUD code hashes and arena bounds. No listener waveform or controller trace; does not identify the original operation/phase selection. Existing initial FIFO replay still applies; do not delete features, change clock or spin-retry. See new trace details.\n\nPreexisting first-DMA trace comparison (2026-10-10): all six Traces (6–11) begin with exactly 2,816 PCM bytes and the same first ten submitted lengths; no gross differing startup rate/epoch observed. Each first accepted submission follows the preceding VI by 7,682–7,726 CP0 Count. The first submitted PCM length divergence from failing Trace 7 is not until write #1,874 for vanilla and #2,184–2,203 among other MKMSZR controls (~36–42s). Fine-grained AI_LEN deadlines diverge earlier and different routes are not matched; equal size histories do not imply identical initial state or musical payload. The first native audio work gate is stock and precedes logos, as documented in Work's draft PR #163; Ghidra additions are in draft #7. Next proposed observer should measure native producer ownership and first later divergence, not duplicate already captured first AI write/length. No additional runtime run, producer code edit or production change. Canonical investigation.\n\nVI-relative audio-service check (2026-10-09): no multi-frame native audio callback gap in failed recordings; upstream FIFO phase remains unresolved. Five independent full observer captures (7–11) record 59,484 first AI_STATUS audio services, one in each active VI; the first 294/295 VI events precede output (plus one unmatched final VI callback in each capture). There are no gaps between active-service VIs. VI→service first-status guest CP0 Count medians stay 7,652–7,722, observed extrema 7,294–9,490, versus 811,092 Count per VI, including in a 352-rejection sustained burst. This disfavors Inventory-triggered whole-frame audio-service starvation as a necessary mechanism; the unresolved variable is AI DMA tail size/completion phase relative to consistently scheduled service. Original scheduler code and Ghidra metadata already map VI/service/audio/RSP dispatch, so broad static retrace is redundant. No ROM/emulator test or production change. See new cross-capture timing analysis.\n\nIndependent endogenous FIFO replay across Traces 7–11 (2026-10-09): zero discrepancies with NO recorded FIFO state, AI_LEN or PCM size as model inputs. Using only recorded service/write/AI-END event times and stock native size recurrence, reconstructed all 59,992 status reads, 59,489 AI_LEN reads, 58,971 submitted buffer sizes and 513 rejected enqueue decisions exactly. In the final Trace 11 rejection, AI END occurs 18 CP0 Count after first FULL-status read and before diagnostic logger's next read, directly changing feedback/escape conditions. This is complete host-event-driven state consistency, not prediction of the originating event schedule and not an approved game/emulator fix. Strictly read-only reusable script and synthetic tests staged in draft PR #162; standalone archive includes original four trace streams. Release remains blocked until a portable upstream correction and runtime proof. See canonical audio investigation.\n\nFull-record closed-loop FIFO reconstruction (2026-10-09): 59,489/59,489 AI_LEN reads and 58,971/58,971 accepted native buffer sizes match the documented native sizing + pinned RMG AI event-time model across vanilla and Traces 7/9/10/11. All 513 rejected guest enqueues also correspond to expected 352-frame generation. This includes both fault onset and recovery, not just a post-failure 3s window. Critically Trace 11's 79th rejection ended when emulated AI END occurred 18 CP0 Count after the first native FULL-status read, before the diagnostic logger's extra status read and the next AI_LEN feedback; new head feedback became 1,464 bytes and 352-frame rhythm resumed. At the start of severe bursts queued tail finishes ~3,784–4,486 Count after next service; isolated failures have ~30,522–31,714 Count of positive headroom. This establishes a phase-boundary FIFO mechanism, but does not yet explain why the patched composition preferentially enters it or prove zero other memory faults. Guest diagnostic logger has a timing cost; emulator fractional DMA v03 may increase FIFO occupancy, so cannot be pre-approved as a fix. No production change; release remains BLOCKED. See full investigation.\n\nTrace (11) overturns the selector-s0 ABI fix hypothesis (2026-10-09): audible acceleration still occurred with the sound call NOP and helper return moved s0→t9. Exact register-only proof SHA 73d8e64b078ff6211b9e059e6f46bdb90f2bef9bd829c2a7e52f41392b77fc68 captured 80 native FIFO rejected submissions over 247.770s: 1 isolated at t=162.671 and 79 consecutive between t=240.087..245.288s, a ~5.2s audible episode that recovered; no long fixed-368 precursor. 14,858 paired VI/memory observations show no out-of-bounds arena or mutated resident-code hashes. 22 separate SDL drops (19,312B) also occur. First attempt emitted no trace when RMG unexpectedly closed entirely after clicking its normal Shutdown button; treat this as unexplained host exit, not an observed guest crash. With muted sound and t9 return, the helper has no memory write and leaves selector downstream argument/used-register invariants intact; phase-sensitive emulated audio timing is the leading inference, but root cause and acceptable portable correction are Pending. Further blind register/cave proofs not justified. Release remains BLOCKED. See selector owner and audio report.\n\nTrace (10) overturns the sound-playback-only hypothesis (2026-10-09): 352 uninterrupted native PCM rejects and audible acceleration despite MOVE SFX wrapper call NOP. Failed MKMSZR_selector-move-call-off_fortress_proof_v01.z64 SHA 7ef2ce16bddcc5716df8aa030c5eaf367363a3bbdf882f6b63ceee2ba73bb25e retains full menu helper but disables sound; in 65.657s, repeated failures span t=42.207..65.607s and user exits while music remains broken. No SDL drops; sampled arena/HUD regions stable. Unlike failing Trace 7, this failure has no long all-368-frame precursor, disproving its necessity. Healthy Trace 9 bypassed whole helper, not only sound. The helper's s0 return-address use is an unverified ABI candidate; a guarded sound-still-off, two-instruction s0→t9 disposable proof (SHA 73d8e64b078ff6211b9e059e6f46bdb90f2bef9bd829c2a7e52f41392b77fc68) is prepared with no production or other feature changes; runtime Pending. Investigate helper effects vs initial AI timing phase before any fix. 1.0 audio gate stays BLOCKED. See selector owner and audio report.\n\nBreakthrough lead, Trace (9), 2026-10-09 — selector MOVE-helper-OFF eliminates the sustained audio failure on a 252-second bounded route; cause not yet proven; 1.0 gate remains BLOCKED. Byte-guarded control differs from failed prefilled Fortress Trace (7) only at Safe Stage Select ROM 0xDD70 (reinstates stock instruction instead of MOVE-SFX helper JAL) and CRC. Audio stayed normal; guest FIFO rejects fell from 74 (73 consecutive) in 130s to 2 isolated over 252s, and longest run of exactly 368-sample buffers fell from 2,383 to 9. Arena highwater, HUD pointer and fixed-code hashes match across tests. This is the strongest single-feature isolation thus far; it does not tell whether actual descriptor 0x1FC SFX, helper register/callback behavior or timing phase caused the change. The staged helper-still-on, native-SFX-call-only NOP ROM MKMSZR_selector-move-call-off_fortress_proof_v01.z64 (SHA 7ef2ce16bddcc5716df8aa030c5eaf367363a3bbdf882f6b63ceee2ba73bb25e) provides next bounded negative control without production changes. Runtime Pending; no emulator automation or other features removed. Stage selector owner and audio investigation.\n\nSafe Stage Select audio lead (2026-10-09): bounded MOVE-SFX-off diagnostic now built, cause Pending. Prior selector confirmation v02 using 0x1FD was rejected after delayed crunchy/choppy gameplay audio; accepted v03a uses only Up/Down MOVE 0x1FC, and its earlier runtime pass did not cover the current intermittent production failure. A disposable SHA-guarded prefilled Fortress proof restores only the selector's original draw instruction at ROM 0xDD70 (plus N64 CRC bytes), leaving every Inventory/Toasty/modern-control/audio-logger patch identical. Output SHA 3988da00cc1fda37d314f4b20452ac678ae17cc1bc2d078c9e0dc12a857dd98f. No runtime executed, no production sound removed; one bounded manual A/B can disprove whether MOVE is necessary, but a clean result could also be timing-phase dependent. See Stage Flow owner and Audio investigation. Release remains BLOCKED.\n\nVanilla Fortress Trace (8) control (2026-10-09): no audible acceleration, but 5 native FIFO-full rejections and 40 host SDL drops; MKMSZR-specific sustained 368-only phase remains unexplained. Clean unmodified USA ROM ran ~321s with RMG v02 observer, entering Fortress at t=53.379s (VI-relative). Four complete streams show an isolated FIFO FULL before Fortress, a three-rejection cluster in Fortress, and another isolated rejection there. The first two head/tail sizes of vanilla's three-event burst match MKMSZR Trace (7)'s sustained 73-event burst exactly (1600/1472, 2816/1472), but at the third reject vanilla uses tail 1408 versus MKMSZR 1472, and recovers. Vanilla has no ~40s all-368-frame regime (longest four consecutive outputs), unlike 2,369/2,383 consecutive outputs in MKMSZR Traces (6)/(7). All 18,978 vanilla accepted writes use the same 551 Count/byte as MKMSZR; 40 host drops (30,944B) occur without reported accelerated music. Thus emulator precision loss/host clipping are shared, but not sufficient to explain mod-only sustained runaway; patch-induced workload/phase interactions and unmonitored corruption remain in scope. The project remains BLOCKED; no ROM/prod change. See full vanilla comparison.\n\nNew audio-clock root-cause candidate (2026-10-09): repeatable ~40-second precursor and pinned RMG integer DMA-duration truncation; production release still BLOCKED. Independent Trace (6)/(7) share nearly identical 2,369/2,383 consecutive 368-frame FIFO production runs, head AI_LEN 1,344→1,280, ending within 0.36–0.53s of the first reject. Pinned RMG computes AI duration using an early integer division (551 rather than ~551.838 Count/byte); all four trace runs confirm 551. The resulting ~22,080.54 guest AI sample-frame consumption against 22,047Hz SDL playback predicts ~134.18 B/s host queue buildup, matching the measured clipping precursor. Native frame builder code matches clean ROM. The guest FIFO trigger remains strong inference, not proven emulator-only cause: quantized feedback and dispatcher timing still require controlled testing. Separate draft diagnostic PR #161 offers a strictly opt-in rational DMA duration A/B condition for exact pinned RMG, with no production or ROM modifications, awaiting CI/runtime validation. See canonical investigation.\n\nProduction rich Inventory music gate (2026-10-09): FIFO recovery/runaway feedback bifurcation now reconstructed; original trigger Pending, PR #156 draft/unmerged. The new exact-v02 diagnostic pair proves 135 additional rejected PCM submissions while audio/VI services each advance 4,934. Synthesis exceeds its nominal interval by 47,616 samples, close to the minimum 47,520 samples represented by rejected buffers. This strengthens PCM omission as a tempo-drift mechanism but does not identify the Inventory trigger, event timing or complete audible loss budget. Preserve the accepted rich presentation. The AI enqueue retry is explicitly rejected. The maintainer ran that populated-Fortress ROM successfully through a long idle/no-audio-defect test, then reproduced tempo acceleration during a rapid Box 1 cursor sweep. New untruncated Trace (5) captures 80 full-FIFO rejects in ~5.27 seconds, no host SDL drops, on-time VI/AI events, and ~24% excess synthesis over accepted PCM during the burst. Follow-up complete Trace (6) and ordinary-items control did not reproduce acceleration: the 219s full sweep had only four isolated native rejects but 14 SDL queue-threshold drops (14,920 bytes), while the 110s ordinary sweep had neither. Both host audio queues grew ~133 bytes/s before overflow, matching the nominal 368-sample/VI producer versus 22,047Hz playback difference; this is independent evidence for host clipping but does not identify the transient guest failure trigger. Cursor/portrait DMA involvement remains a hypothesis because input and ROM DMA were not logged; no production correction was built. A focused comparison of first failures in Trace (5) versus all four isolated Trace (6) failures now identifies the next-service FIFO-state fork: a 1,472-byte queued tail in Trace (5) persists 4,222 Count beyond the next service, making the maximum 704-frame recovery buffer go into a still-busy FIFO and yielding AI_LEN=0; a 1,408-byte tail in Trace (6) drains ~30,628 Count beforehand, so the same 704-frame buffer becomes the new head, AI_LEN reads 2,808 and the next synthesis shrinks to 352. All AI durations are 551 Count/byte and VI is 811,092 Count, only 20 Count longer than a nominal 368-frame DMA. This explains intermittent amplification/recovery, but not the upstream reason the first FIFO/tail state differs. See upstream report, source evidence and bounded matched test.
Cross-feature corruption and memory audit (2026-10-09): no fixed 16 KiB expansion-payload/arena-floor overlap identified; dynamic headroom Pending. The historically recurrent accelerated/crunchy-audio symptom has multiple previously proven non-audio causes; do not confine the root-cause search to rich HUD or blame the emulator. On the exact instrumented Fortress ROM, file 1A occupies [0x801AF820,0x801B3420) and stops exactly at the guarded arena floor. Named production runtime suballocations and PR #156/diagnostic proof ranges are disjoint; selected former bad sites (false-zero dispatch cave, GAME SETTINGS stale store and HUD first-load size) remain corrected. Earlier all-stage 16 KiB proof cursor left 30,120 bytes, but the newer accelerated post-Kia Fortress snapshot left only 17,496 bytes relative to the same static bump boundary: incomparable compositions, not proven exhaustion, but peak allocator and rewind/lifetime instrumentation are the next priority. The 16 KiB reservation, Toasty, and modern controls are not rejected. See broader audit and Memory Map.\n\nFull-project rich Inventory/audio blocker (2026-10-09): cross-feature memory-provenance diagnostic now staged; root cause and production guard still Pending. The audio FIFO feedback runaway/recovery mechanism is trace-confirmed but may be a symptom of any bad composed feature. The 16 KiB fixed allocation is non-overlapping; dynamic peak/rewind/stale-pointer safety is not exhaustively certified. Draft PR #160, based on separate host observer #157, adds opt-in VI-sampled arena/HUD pointer and code-digest metadata without touching ROM. Synthetic C/injection tests passed; one exact v02 pinned Flatpak CI build passed on commit f6c67e5 (Actions run 37899907006), while observer-off parity and manual runtime proof remain Pending. New size precursor evidence: Trace (5) contains three 400-frame buffers in its last 150ms before the burst, compared with one before Trace (6)'s isolated failure. No fixes or feature deletions authorized. The project release gate remains BLOCKED. See primary investigation.\n\nTrace (7) v02 memory/provenance result (2026-10-09): accelerated-music failure reproduced; no monitored arena/code/HUD invariant violation; release gate BLOCKED. Maintainer's 130-second prefilled Fortress route captures 74 paired native FIFO rejection events (one isolated at t=82.112 and a 73-rejection burst t=118.729..123.529), with zero lost records. The v02 observer samples arena high-water 0x8028CD48 (15,432 B below 0x80290990); resident cursor 0x8028C538, HUD data pointer 0x801FEBA0, and three ROM-matching code hashes stay unchanged through both failures. Last key-portrait-cache change precedes the severe burst by >63 seconds. Recurrent FIFO feedback bifurcation is stronger; exhaustive memory/legacy-patch corruption is not excluded by VI sampling. Separate SDL host discards occur both before and during failure. Continue upstream cadence/source review, not feature removal or FIFO retry. See primary audio investigation.\n\nNative Inventory legend (2026-10-06): Runtime-confirmed bounded. The maintainer accepted the focused legend v01 proof as working perfectly. The native Inventory panel now shows dynamic BOX n OF 4; Items mode shows LEFT-RIGHT BOX, A USE, B COMBINE, R POWERS; Power Ups uses the approved sparse layout L-R POWER and R ITEMS. The patch replaces only the stock legend text-emission block and appends after the already Runtime-confirmed in-Inventory switch module; the panel background and switching semantics remain unchanged. The separately investigated Inventory/loading hang is unresolved and is not attributed to or bundled with this legend promotion.
Powers-as-pickups OFF earned-XP persistence correction (2026-10-06): Runtime-confirmed bounded. The prior OFF composition correctly retained stock XP awards/caps but did not synchronize native current XP into MKSV persistent XP or install the safe stage-entry restore, so stage exit/re-entry could drop earned powers. PR #154 keeps stock XP behavior untouched, snapshots current XP into existing state +0x44 at the established lifecycle-v06 teardown seams, and reuses the Diagnostic-B stage-init restore before four-box reconstruction. The maintainer manually validated the focused all-stage proof ROM MKMSZR_earned-xp-persistence_all-stages_proof_v01.z64 (SHA-256 05c579ad5473a3f357b4cae45b660dbcbdd5b5aa66389eff3395a8e47bf8a550) and reported the requested earned-XP/power persistence route worked. The fix adds no allocation, does not call the unsafe tier evaluator at stage init, and leaves Powers-as-pickups ON unchanged.
In-Inventory four-box switching (2026-10-06): Runtime-confirmed bounded in full production composition. Disposable production v03 used current main's real global materializer, lifecycle v06, complete Inventory HUD, GAME SETTINGS/modern controls, gameplay box shortcut, BOX n OF 4, Fire/Bridge 40-slot ownership correction, and donor-backed Toasty/Temple-audio composition, then added the audited Inventory-loop switch path. The maintainer reports the result works perfectly. In Items mode, Left/Right now cycle all four authoritative boxes with wraparound while Inventory stays open; empty boxes remain navigable, pending Combine is canceled, selection resets to the first occupied row, foreign credentials remain stage-masked, and Power Ups retain stock horizontal navigation. The existing gameplay Block + Use + Left/Right shortcut remains intact. Proof ROM SHA-256 14feee5d6a0a79a20359781364eb6a9ef77ffffad3f2c0bee94760b9c2074f7e. Runtime scope is bounded; exhaustive lifecycle/full-seed coverage remains part of the final release gate.
Fire/Bridge multi-box credential gate correction (2026-10-06): Runtime-confirmed bounded. Stock handlers 0x8007206C (Fire IDs 0x17..0x19) and 0x80071FE8 (Bridge IDs 0x1D..0x1F) require all three credentials but stock scans only the ten-slot LIVE inventory. The production correction changes only each handler's ownership scan base/count from LIVE 0x800A600C, 10 words to authoritative backing 0x800A6048, 40 words; native physical use-site gating and completion actions remain stock. Disposable v03 composed the real production four-box switching/masking path and BOX n OF 4, distributed each three-credential family across multiple boxes, and used proof-only placement/gate bypass at the exact Fire/Bridge use-zone trigger records. The maintainer manually switched boxes and confirmed both positive and negative controls in both Fire and Bridge: all three owned across boxes succeeds; one credential genuinely absent fails. Positive SHA-256 9952a2ceb4d3e64888bb1f2662560a703f15db0613fd07043906a8ca6539be25; negative SHA-256 f6521862d51ab2899184e41277e56e11c0d2f43630b732ff5e7996218236d812. Runtime scope is the 40-slot ownership correction; the proof bypassed the unchanged stock physical use-site gate for convenience.
Inventory HUD production wording follow-up (2026-10-05): Runtime-confirmed closed. First webapp runtime after #146/#147 exposed that #146 had promoted proof-only slash wording into production. Plain / is a known-bad native-font glyph. PR #148 restored the earlier accepted REQ. POWERS X-Y + right-aligned CHKS XX-85 (Vanilla: REQ. XP 5100), and the maintainer subsequently generated and tested a deployed-webapp ROM and reported that it looks correct. The regression reproduced with Powers-as-pickups both ON and OFF and was presentation-only; Required Powers remains independent of pickup mode.
Native Inventory HUD closure (2026-10-05): v20 Runtime-confirmed the first-load 0x1200 allocation correction; v21 then Runtime-confirmed the completed TEST LAB Inventory presentation with all nine Power pickups, colored STG# credential labels, real paper titles, streamed portraits, ordinary-row font/palette forwarding, plus the accepted native-font-safe status format (REQ. POWERS X-Y, CHKS XX-85); #146 temporarily regressed those strings to slash forms and the follow-up restores them. A dense debug layout exposed a separate rapid-Power lifecycle edge: v22 proved the white-flash wrapper's synchronous sleeps were causal, and v23 Runtime-confirms the accepted correction: award synchronously, run the three 2-tick white pulses in one guarded child process, and allow later rapid Power pickups to award immediately. The shared browser/CLI integration now promotes this architecture without adding a new flash allocation. Final representative full-seed production-composition runtime validation remains the release gate.
Focused lifecycle update (2026-10-01): lifecycle v06 is Runtime-confirmed and promoted to the shared production pipeline. The exact accepted reference is Very Hard / 9 total lives / 5 continues / persistent HP ON. The shared core now parameterizes that same lifecycle control flow for build-time difficulty, starting lives (1..10), starting continues (0..5), and persistent HP ON/OFF; browser defaults are Very Hard / 5 / 3 / HP ON. Non-reference combinations are Implementation/CI-confirmed pending representative runtime coverage. v04 remains Rejected / unsafe composition. See Lifecycle static closure, Persistence/inventory/lifecycle, and Web patcher/product behavior.
2026-10-27 Wind/XP/debug correction — Runtime-confirmed bounded; promotion approved. The maintainer accepted MKMSZR_keys-xp-debug_all-stages_proof_v01.z64 (SHA-256 7129c455a29f736c77bb06b6a34d5695326f6eae48981b2a82a8739e47bc8465, CRC 4D2B66FA / 89A89E6C) with “It's good, merge it.” This closes the bounded manual gate for PR #174's Wind portraits, OFF stage-XP limits and debug starting/reset inventory. No separate per-step log or exhaustive stage/seed matrix was supplied. Proof configuration: ICONXPDEBUG01, Powers as pickups OFF, debug keys ON, one life, zero continues. Static tests cover both modes and reset store bounds; the assistant ran no game emulator. Reproduction and the proposed manual route remain in docs/key-icons-xp-debug-validation.md. Existing audio/release gates remain unchanged.
Long v05 stress control 2026-10-10 — Runtime-confirmed no sustained music speedup, short crunches observed; trace-confirmed recoverable native/host losses. 598 seconds anchored to first PIF reply (first VI missing due bounded ring wrap), 9,766 valid player PIF reads, 348 Down and 264 Up rising edges, intensive cross-box/combative activity user-reported; 18 unique native guest PCM failures (not 36 duplicate FULL logger reads), largest 5 over 0.266s, at least 59 SDL threshold drops retained (first 11,422 host events overwritten). Input/AI/RSP complete; VI/phase lost 12,668/2,877 earliest samples. After last pressed input t430.829, 167.38s of neutral input still include 10 guest rejects and 23 retained host drops. Three 38.5–44.8-second 368-frame plateaus each step 35,000 Count. Two ~40s healthy 368-only stretches show AI END mean 811,077.40/811,077.75 while native services remain ~811,092, accounting for quantized AI_LEN drift. Slow case's next service is +35,580 Count after promoted 1,408B tail deadline (352-only + median 35,226 Count FIFO idle); fast case is −4,126 Count before 1,472B tail deadline (repeated rejected PCM). Rejected events were separated by successful services; no all-368 causal precursor established. Native producer pointers/parameters stable, earliest event timing perturbation still Pending. See canonical investigation.\n\n2026-10-10 native rich Inventory audio: three fresh RMG v04 captures — maintainer Runtime-confirmed clean, slowed/clipped, and sped-up/clipped sound; read-only host traces independently confirm 352-frame undersupply and 81-reject oversupply. Five streams each, all saved=total, overwritten=0. Clean 379.61s (5 guest rejects, no major audible symptom), slow 229.80s (5 isolated rejects; 2,303 consecutive 352-frame accepted outputs following one rejected service (2,304 accepted 352-frame buffers in the wider interval), −4.35% native synth frames/VI over last 38.4s, host queue median 504B), fast 277.60s (81 recurring guest FULL rejects (~one per four audio services; intervening accepted submissions) in 5.33s, producer sample time +23.8% over target, host queue median 16,600B). Slow loop arises after a DMA END 10 CP0 Count following first native FULL-status read. DAC/manager params and all AudioInfo/command/synth addresses stable at 52,316 phase snapshots; 52,313 synth sample-clock increments match expected sizes exactly; fixed code hashes stable in Fortress. The healthy run also has 39.77s all-368 output, refuting this as a sufficient acceleration indicator. No ROM identity embedded in traces; startup route and cursor trigger remain user report, not logged controller events. Neither original upstream operation nor release-ready fix identified. Canonical investigation.\n\nProduction rich Inventory (2026-10-09): cursor-sweep PCM failure Runtime-confirmed bounded; upstream trigger and root-cause correction Pending. Original environment is Bazaar-installed RMG Flatpak with ParaLLEl RDP + ParaLLEl RSP; exact versions/audio settings are not recovered. AI_LEN 1344→1288; one 2,310-buffer plateau has zero pressed buttons. AI_LEN=1216 29 times and accepted 1,600B PCM 30 times without runaway. Retained VI each receives audio service, sample clock consistent; Grunt AI/RNG state unobserved, not disproven or implicated. Canonical owner.\n\nRMG v05 input-aware accelerated Inventory capture (2026-10-10): maintainer Runtime-observed speedup/clipping within ~one minute of Inventory use; trace confirms exact chronological PIF input / audio correlation, not initiating game cause. 1,594 PIF responses, 3,962 native complete audio services (no active-VI gaps), 86 guest PCM rejects over 5.734s following 73 Up/Down button presses in the 22.8s preceding rejection. Probable Inventory opening Z at t38.786s; first repeated navigation t39.711s; first AI_LEN=1,216 t62.093997 precedes nearest new Up response by 0.535ms, first 1,600B at t62.111, first FULL at t62.544. Promoted 1,472B DMA remains unfinished 4,056 Count beyond next service; prior v04 fast case was −4,126 Count. 86 failed vs 259 successful submissions between first/last rejections, +23.54% synthesized sample frames relative to target, zero SDL queue-threshold drops. Last PIF button release t63.395; failures persist until t68.278 with neutral input. Low AI_LEN 8B staircase descends about every 4.804s after Inventory activity; cause may be opening/menu versus cursor/audio work and remains Pending. All six trace rings complete, native sizing/clock 0 mismatches, no ROM identity embedded. Full canonical details in Inventory upstream investigation.\n\n2026-10-10 v04 deadline accounting — Trace-confirmed and statically consistent, production fix Pending. Normal 368-frame output at 551 Count/PCM-byte takes 811,072 Count, just 20 fewer than the observed 811,092 Count/VI; 352 and 384 neighboring frame sizes differ by ±test_musicnormal.state has one cumulative failure; test_musicspeedup.state has 136. The first failure predates the healthy capture; the last is seven WESS ticks before accelerated capture. The accepted proof remains bounded success for a different composition. Audio and VI deltas match; upstream chronology is missing. The guarded data-only pre-filled Fortress ROM entered Inventory and completed the maintainer's full idle test (plus extra minute) without audible defects, then a separate cold-boot rapid cursor-sweep route reproduced tempo acceleration. Four zero-overwrite Trace (5) streams establish 80 paired native FIFO enqueue rejects between t=42.357..47.624, 80 reconstructed rejected 352-sample buffers (28,160 omitted frames), a self-terminating ~24% extra synthesized relative to accepted sample time, nominal VI/AI deadlines, and no host SDL drops. A subsequent 110-second ordinary-items-only test and 219-second full sweep did not reproduce acceleration. The ordinary test recorded zero AI FIFO failures or SDL drops; full sweep recorded four isolated guest FIFO rejections and 14 host queue-limit discards (14,920 bytes). Both host audio queues grew ~133 bytes/s during the pre-drop interval, consistent with 368×60 stereo samples/s produced versus 22,047Hz consumed. This supports a separate host-clipping mechanism and weakens deterministic cursor/portrait causation; actual per-item display validation, exact cursor operation and guest upstream trigger remain Pending. Trace (4) files are hash-identical to prior Trace (3), not independent evidence. No emulator automation or enqueue retry was performed. PR #156 stays draft/unmerged. See paired report and one bounded matched manual experiment.
Cross-feature reservation audit (2026-10-09), ROM-static and historical runtime evidence only; peak safety Pending. Exact clean/instrumented/prefilled ROM identities verified; raw file-1A transport fills but does not cross [0x801AF820,0x801B3420) with unchanged floor. Live false-zero stock action records stay byte-identical; GAME SETTINGS trailing bad store remains suppressed; rich HUD first-load size set correctly in the current generated layout. All named production fixed RDRAM suballocations are disjoint; draft #156 and diagnostic tail are separate proof-only owners. The old 16 KiB all-stage proof had 30,120 B peak headroom, while the later failing Fortress endpoint has 17,496 B immediate headroom; these are neither matched peaks nor direct evidence of allocator overflow. Other features and cross-composition faults remain in scope; no automatic feature removal or ROM correction.\n\nTrace (5)/(6) feedback reconstruction (2026-10-09), Static/trace-confirmed mechanism; upstream trigger Pending. At the first failure in Trace (5), the 1,472-byte queued tail finishes 4,222 Count after the next audio service; in Trace (6), the 1,408-byte queued tail finishes 30,628 Count before it. Both first failure services have nearly identical VI-relative Count offsets (7,592/7,596). The differing next FIFO occupancy flips feedback for an accepted 704-frame buffer: AI_LEN=0 under a queued tail in Trace (5) versus AI_LEN=2,808 when newly started as a head in Trace (6). All four Trace (6) failures recover through an empty FIFO at the next service. The Trace (5) sustained burst ultimately stops when a head ends 736 Count before service. DMA timer is 551 Count/byte and VI=811,092 Count, just 20 Count above nominal 368-frame DMA. No Inventory cursor/PI event or task-generation provenance is logged; this is not an Inventory-origin or universal backend defect claim. See upstream investigation.\n\nOpt-in RMG memory probe v02 (2026-10-09): implementation/static smoke only; RMG/ROM runtime Pending. Draft PR #160 extends #157, not #156, by sampling guest physical RDRAM without ROM changes or guest AI MMIO. Synthetic C read-only/idempotence checks passed and exact pinned v02 Flatpak artifact built/linked in Actions run 37899907006 from commit f6c67e5; guest runtime and opt-out parity remain Pending. The tool can report arena cursor > upper bound, a suspicious cached HUD pointer, and changes in selected fixed-code digests during a stable stage, but zero VI-sampled warnings cannot rule out transient between-frame corruption. Exact initial upstream writer or allocation defect remains Pending, and production is blocked.\n\nVI/audio service offset audit (2026-10-09), Trace-confirmed; no runtime correction: Across Traces 7–11, 59,484 native first AI_STATUS reads map uniquely to 59,484 active VI events, after 294/295 initial VI events before audio starts, plus one final unmatched VI callback in each recording; there are no gaps between active-service VIs. VI→first-status CP0 offset medians are 7,652–7,722 Count (range 7,294–9,490), versus normal 811,092 Count/VI, including Trace 10's 352 consecutive rejected submissions. This does not support missing or duplicate audio service calls or whole-frame scheduler starvation during the observed failures, but doesn't identify why AI DMA deadlines/queued-tail sizes become unfavorable. See canonical investigation.\n\nStronger event-only FIFO replay (2026-10-09) — trace/static-confirmed, correction Pending. On the five complete RMG captures, independently reconstructs 59,992/59,992 AI FIFO status reads, 59,489/59,489 AI_LEN reads, 58,971/58,971 written PCM sizes and 513/513 guest-native rejected decisions. Uses only previously observed audio service/start/completion timestamps plus native size formula as inputs; all FIFO contents, register values, accepted/rejected outcomes and sizes are calculated internally and checked against captures. Acknowledges the observed timing schedule is exogenous and diagnostic logger can change timing. No emulator/runtime correction, ROM patch or memory-safety proof is inferred. Reusable independent script staged in draft diagnostic PR #162; see full audit.\n\nFull five-capture native audio closed-loop audit (2026-10-09): Trace-confirmed exact feedback/size, upstream trigger Pending. Across sound-on Trace 7, clean vanilla Trace 8, entire-helper-off Trace 9, muted-helper Trace 10 and s0-preserved muted-helper Trace 11: all 59,489 AI_LEN reads exactly equal pinned RMG FIFO head/deadline/Count reconstruction; all 58,971 native accepted PCM buffer sizes exactly equal the prior AI_LEN size recurrence; all 513 native rejects calculate to 352 synthesized sample-frames. Severe burst first-reject queued tail is still active at next service (−3,784..−4,486 Count), while isolated full reads have already drained (+~30,500..31,700 Count). Trace 11's final rejection coincides with an emulated AI END 18 Count after the first FULL status read and before the second logger-only read, immediately changing the following AI_LEN to 1,464B and allowing recovery. End-to-end audio event accounting, not a runtime-validated production correction or proof of general guest-memory integrity. Diagnostic logger changes instruction timing. See focused full audit.\n\nSafe Stage Select helper active, SFX call OFF, s0-preserved Trace (11) — 2026-10-09; Runtime-confirmed transient accelerated music, trace-confirmed FIFO burst, no fix. Exact two-instruction register-only proof SHA 73d8e64b078ff6211b9e059e6f46bdb90f2bef9bd829c2a7e52f41392b77fc68 captured 247.770s zero-overwrite with 80 rejected native submissions: one isolated at t=162.671s, then 79 consecutive at t=240.087..245.288s, automatic recovery. No 40-second 368-only precursor (maximum 7). Stage-9 arena cursor/HUD/hash checks remain bounded/stable; separate 22 SDL drops total 19,312B. First attempt heard prolonged distortion, then RMG's normal Shutdown unexpectedly terminated the whole GUI without any trace export (observer dumps on process atexit; mechanism unknown). This disproves the s0 ABI encoding change as a sufficient cure; a distinct phase-dependent helper/core timing explanation remains hypothesis and production audio gate blocked. See Stage-Flow and canonical audio analysis.\n\nSafe Stage Select helper ON / SFX call NOP Trace (10), 2026-10-09 — Runtime-confirmed accelerated and clipped music; 352 consecutive FIFO failures. 65.657s clean four-stream RMG v02 capture from the static-verified Fortress proof SHA 7ef2ce16bddcc5716df8aa030c5eaf367363a3bbdf882f6b63ceee2ba73bb25e, whose only change from failing Trace (7) was removing JAL 0x80064C18 at helper ROM 0xE29C and CRC. The first native failure at t=42.207s was followed by one per native service until t=65.607s when the user quit; 352 rejected guest submissions, no host SDL queue drops. Accepted buffer sizes cycled 688/416/368 frames, without Trace (7)'s long 368-only precursor; arena/HUD/FNV memory samples unchanged. Thus the native movement SFX call is not required for prolonged acceleration under these conditions. A two-instruction callee-saved-register isolation proof (return-address holder s0→t9, SFX still off) was statically built for one bounded follow-up; actual register corruption not proven and runtime pending. See selector owner and audio report.\n\nSelector MOVE SFX-off Fortress Trace (9), 2026-10-09 — bounded Runtime-confirmed healthy audio, trace-confirmed disappearance of sustained FIFO/368-only regime. Exact prefilled Fortress Trace (7) ROM was changed only at ROM 0xDD70: JAL movement-helper → stock li a0,0x50 plus N64 CRC. A 251.722s cold-boot observer capture had no perceived acceleration, only two isolated guest FIFO-reject attempts, and a maximum 9 consecutive 368-sample buffers. Earlier SOUND-ON Trace (7) had 74 rejections including a 73-failure burst and 2,383 consecutive 368 buffers within 130.062s. RDRAM reset/base/cursor bounds, cached HUD pointer, and all three fixed-code hashes remained unchanged. The 22 separate host queue drops do not imply speed-up. This strongly implicates the selector helper/SFX path or the audio phase it changes, not a proved invalid memory cave; a single benign test is not causation. Narrow second proof retaining the helper but NOPing only JAL 0x80064C18 at 0xE29C was statically built; runtime Pending. Release gate remains BLOCKED pending repeated sound-call-vs-helper isolation and correction validation. Selector owner.\n\nVanilla MKMSZ Fortress negative control / Trace (8), 2026-10-09 — bounded Runtime-confirmed absence of audible acceleration; native FIFO FULL trace-confirmed. The clean USA Rev.0 ROM ran 321.460s from first VI and reached stock stage 9 at t=53.379s. Complete zero-overwrite observer streams show five actual guest FIFO FULL events (one before Fortress, three consecutive at 124.558..124.691s, one later) and 40 SDL drops totaling 30,944B, despite no perceived acceleration. Vanilla's third failure uses queued tail 1,408B and recovers; MKMSZR Trace (7) retains 1,472B and suffers 73 repeated failures. Vanilla longest 368-only run is 4 vs MKMSZR 2,369/2,383, though all use the same emulator 551 Count/byte. Stock guest-memory probes do not implement MKMSZR marker semantics. This is evidence for a shared audio timing vulnerability and a composition-specific sustained trigger; no single ROM fault, emulator-only resolution or production guard confirmed. See Trace (8) comparison.\n\nCross-capture AI clock analysis (2026-10-09), Trace-confirmed/Static-confirmed, production fix Pending. Two independent prefilled-Fortress recordings contain 2,369 (Trace 6) and 2,383 (Trace 7) consecutive 368-frame native submissions over nearly 40 seconds with head AI_LEN 1,344→1,280 before a FIFO FULL. Across four runs, all accepted emulated AI durations have exactly 551 Count/byte; pinned core's pre-multiply integer truncation is statically verified and predicts ~134B/s host SDL queue buildup, matching prior observed ~132–133B/s. However a fractional-duration control has not been runtime-validated and this clock discrepancy alone does not certify absence of unrelated memory corruption. Draft emulator-only PR #161 is Pending; no game ROM modification permitted by this result.\n\nRMG v02 opt-in memory trace (7), 2026-10-09 — Runtime-confirmed symptom, bounded memory negative. Four complete streams from prefilled Fortress record 7,798 VI+memory sample pairs, one native FIFO-full rejection at t=82.112s, then 73 between t=118.729 and 123.529s; host-side SDL drops 4 (2,832 bytes), no overwritten records. Arena cursor remained <=0x8028CD48 (15,432 B of sampled headroom), HUD pointer 0x801FEBA0 remained within current cursor, and all three observed rich-HUD/legend/materialized fixed-runtime checksums exactly match the prefilled-ROM source after initial load. The lasting cursor value 0x8028C538 and key-portrait cache were unchanged near the burst. A once-per-VI probe cannot certify absence of transient or unmonitored corruption. This confirms the new observer records the failure, not a production correction; root cause and gate Pending. See full reconstruction.\n\nNative Inventory legend (2026-10-06): Runtime-confirmed bounded. The maintainer manually accepted focused legend v01 and reported it works perfectly. The proof preserves the production in-Inventory switch path and native panel, adds dynamic BOX n OF 4, uses LEFT-RIGHT BOX / A USE / B COMBINE / R POWERS in Items mode, and the intentionally sparse L-R POWER / R ITEMS layout in Power Ups. This evidence is presentation-only and does not resolve the separately reproduced Inventory/loading hang.
In-Inventory four-box switching (2026-10-06): Runtime-confirmed bounded, full production composition. The maintainer reports production-composition v03 works perfectly. Items-mode Left/Right cycles all four authoritative boxes with wraparound without closing Inventory; empty-box traversal is healthy, selected-row state refreshes coherently, stage masking/SEALED behavior remains intact, and the existing gameplay box shortcut and Inventory HUD coexist. Power Ups retains stock horizontal navigation. Proof SHA-256 14feee5d6a0a79a20359781364eb6a9ef77ffffad3f2c0bee94760b9c2074f7e. This does not replace the final representative full-seed/lifecycle release gate.
Fire/Bridge four-box credential gates (2026-10-06): Runtime-confirmed bounded. Disposable v03 used the real production four-box switch/mask implementation and visible BOX n OF 4, with Fire IDs 0x17..0x19 and Bridge IDs 0x1D..0x1F distributed across separate authoritative boxes. The maintainer manually switched boxes and reported both stages working perfectly in paired positive/negative tests: all three owned across the 40-slot backing inventory succeeds; the otherwise-matched build with the third credential absent fails. Positive SHA-256 9952a2ceb4d3e64888bb1f2662560a703f15db0613fd07043906a8ca6539be25; negative SHA-256 f6521862d51ab2899184e41277e56e11c0d2f43630b732ff5e7996218236d812. For test speed, v03 placed the player from the actual Fire/Bridge use-zone trigger records and bypassed the unchanged native physical use-site gate; therefore the runtime-confirmed claim is specifically the corrected 40-slot ownership scan and box composition, not a revalidation of the stock trigger predicate.
Powers-as-pickups OFF earned-XP persistence (2026-10-06): Runtime-confirmed bounded. PR #154's focused all-stage proof MKMSZR_earned-xp-persistence_all-stages_proof_v01.z64 (SHA-256 05c579ad5473a3f357b4cae45b660dbcbdd5b5aa66389eff3395a8e47bf8a550) retained stock XP acquisition and used the OFF-only MKSV +0x44 snapshot/restore path. The maintainer manually ran the requested earned-XP/power re-entry validation and reported that it worked. Scope is the tested OFF lifecycle route; this does not replace the final representative full global-seed release gate.
Inventory HUD production wording regression (2026-10-05): Runtime-confirmed closed. A ROM generated through the deployed webapp after #146/#147 rendered the slash separators in the promoted REQUIRED POWERS X/Y / XX/85 strings as corrupted native-font glyphs; the maintainer reproduced the issue with Powers-as-pickups OFF and ON. PR #148 restored REQ. POWERS X-Y, CHKS XX-85, and Vanilla REQ. XP 5100, with matching runtime digit offsets. After deployment, the maintainer generated a new webapp ROM and reported that the corrected Inventory display looks good. This closure is bounded to that deployed production route and does not change Required-Powers semantics.
Inventory-HUD TEST LAB line (2026-10-05): v16/v18/v19 establish the preexisting Inventory hang and reject later pickup-manager/resource count as its owner. v20 Runtime-confirms the one-word first-load size correction removes that hang. v21 Runtime-confirms the completed nine-Power TEST LAB Inventory presentation and ordinary-row fifth-argument correction. The exact v21 artifact used slash-form status strings that are retained only as proof history; deployed production wording is now Runtime-confirmed after PR #148 as REQ. POWERS X-Y + CHKS XX-85. Final all-stage/full-seed production-composition coverage remains Pending. See detailed diagnosis and Native HUD and UI.
Earth hazard discriminator (2026-10-05): the photographed horizontal wall spikes in Earth were inactive at difficulty 2 and became active in an otherwise identical proof after changing only the difficulty halfword to 4 (Very Hard). This is Runtime-confirmed for that bounded 2→4 comparison; it is not a claim about every Earth hazard or every difficulty value.
Power Upgrade presentation update (2026-10-05): the original Ice Blue Herbs and three-pulse white Sub-Zero presentation remain Runtime-confirmed. Dense v21 placement exposed a rapid-pickup lock when the pulse loop slept inside the pickup callback. v22 direct-award control removed the failure. v23 Runtime-confirms the accepted nonblocking architecture: progression awards synchronously; one guarded child owns the same three 2-tick white pulses; rapid later Power pickups continue to award. Production preserves the existing flash allocation; final full-product composition runtime remains Pending.
Lifecycle update (2026-10-01): lifecycle v06 is Runtime-confirmed on the bounded manual route and promoted to production. The accepted build preserved damaged HP across Pause -> Quit -> re-entry, preserved lives/continues/XP/powers/four-box inventory, allowed repeated ordinary deaths below the former ×7 floor with full replacement-life HP, preserved run state across accepted Continue, and reset run progress on genuine final Game Over to Very Hard / 9 total lives / 5 continues / starter inventory while retaining all five GAME SETTINGS preferences. No music-speed corruption was reported on the accepted v06 route. v04 remains Rejected / unsafe composition.
-
Implementation/CI-confirmed: PR #129 implements the researched Wind/Water/Earth/Fire/Prison destination actions, generic Bridge/Fortress activation gates, portable Strength award separation, and Prison credential reconstruction after the stock reset. Current-main lifecycle v06 is unchanged. CI run #1583 passed.
-
Static-built proof:
MKMSZR_destination-wrappers_first-item-swap-proof_v04.z64, SHA-256c7df1ca9798d6532783a80205b123a11e5c7655cda069342e6c60b4ab75547ee, CRC1/CRC259037DDA / 4BACE8C6. -
Allocation correction: the earlier high-tail materializer placement
0x801B2960..0x801B2EBF/0x00F6B140..0x00F6B69Fis Rejected after a real web build exposed overlap with fixed Toasty audio. Corrected helper runtime is0x801B2310..0x801B28EF, ROM0x00F6AAF0..0x00F6B0CF; lifecycle v06 remains0x801B0A60..0x801B0FF3. Wrapper behavior was not changed by the relocation. -
Proof scope: stock reward visuals isolate wrapper/state behavior; foreign-resource visual conversion is not re-tested. Prison uses proof-only backing-box mirroring after key award so the exact entry reconstruction helper can be tested without the full four-box frontend.
-
Runtime status: Runtime-confirmed bounded. The user manually validated v04 and reported the tested first-item/key swaps and location-owned checkpoint behavior working perfectly. This confirms the exercised logical-reward-vs-destination-state separation, including moved keys awarding without carrying their source checkpoint and original key locations retaining their physical checkpoint/state behavior. It does not claim exhaustive coverage of every item/location permutation.
-
Proof packaging correction (2026-10-02): v01 is superseded as a validation package because it omitted Safe Stage Select. v02 added the Runtime-confirmed compact eight-stage selector presentation plus its selector-specific one-shot auto-save bypass. v03 clustering is also superseded by the cleaner first-item-swap v04 design.
-
First-item-swap v04 validation correction (2026-10-02): v03's coordinate clustering is superseded. v04 returns to v02's exact wrapper/Safe-Stage-Select base and swaps logical/visual identities only: Wind first Herbs <-> Circle, Water first Mana <-> Triangle, Earth first Formula <-> Square, Fire starting Potion <-> Triangle Up, and Prison first Herbs <-> L1 key. Each moved key therefore awards from a non-checkpoint destination, while its original physical key location holds the mundane first-stage reward but retains the researched destination wrapper. User manual validation reported the proof working perfectly. This directly Runtime-confirms the exercised key-without-checkpoint, checkpoint-without-key, and late revisit/non-downgrade behavior.
Runtime-confirmed bounded Audio-1 isolation. The user confirmed Temple played the seeded “Friendship!” clip while the previously affected Bridge and Prison robot/miniboss audio remained vanilla. The proof preserved stock descriptors 0x41/0x42/0x43, replaced only Temple events 123/124, and routed through isolated 486 -> 329 -> 319. Audio-2 event-125 isolation is also Runtime-confirmed: TEMPLE-0003 -> shao-laugh behaved as expected, with stock opening Scorpion lines and the later selected Audio-2 replacement. Proof SHA-256 f0c87bbc52365361bcbc288f5db37d3918deee85136bef29728229cf4efb8a12.
MKMSZR current documentation · Wiki source: wiki/ · Evidence/archive: ChatGPT Library MKMSZR Research · Product source: smeagol44/MKMSZ-Randomizer
Orientation / Project
- Home
- Contributor start here
- Project status
- 1.0 requirements and roadmap
- Runtime validation status
- Architecture overview
- Research methodology and evidence
- Glossary
Core Technical Reference
- Core runtime and native payload
- Memory and allocation map
- Function registry
- Address and patch-site registry
- Data structures and encodings
- Resource and overlay system
- Address quick reference
- Testing and CI
Randomizer / Product Systems
- Stage flow and selector
- Pickups and stage-local randomization
- Global item materialization and solvability
- Persistence, inventory and lifecycle
- Lifecycle static closure / unbuilt v05
- XP and progression
- Native HUD and UI
- Presentation and branding
- Audio system
- Palette and recoloring
- Web patcher and product
Stage Reference
Gameplay / Enemy / Donor Research
- Enemy randomization
- Player actions and special moves
- Fighter recompiler plan · status
- MKT compatibility overview
- MKT adapter primitives
- MKT fighter asset translation
- Sub-Zero ↔ Sektor animation mapping
- Sektor takeover proof history
- Toasty audio research
- Toasty visual research
- PS1 research
- N64–PS1 comparison
- N64 -> PS1 port planning and strategy
History / Failures / Superseded Material