[M11] Rebaseline real-world compatibility from expanded corpus

[Teakowa]

Goal

Close M11 from current real-world evidence without turning reference-output parity into an open-ended product objective, then hand off to the tooling-first roadmap defined by #88.

Current reality

M11 began as an evidence-first expansion of Wright's real-world .opy compatibility corpus. That evidence has now driven several bounded corrections:

• [M11] Evidence-first real-world compatibility hardening (phase 1) #81 — phase-1 corpus acquisition/classification
• [M11 correctness] Preserve included-file identity for frontend diagnostics #83 — included-file diagnostic/source identity correctness
• [M11 contract] Align deferred .opy surface with corpus evidence #84 — deferred .opy surface documentation
• [M11 evidence] Add settings-free real-world parity candidates #85 — settings-free candidate investigation
• [M11 settings] Typed custom-game-settings support in the native frontend (v1 corpus boundary) #86 — typed corpus-bounded custom-game-settings support
• [M11 correctness] Close supported-surface Workshop emission and round-trip divergences #87 — supported-surface Workshop emission/round-trip closure; implementation is on current main, but final independent QA acceptance is still required

The architecture/product interpretation has also changed materially. #88 is now the authoritative rebaseline task: Wright is tooling-first, owns compiler-grade semantic frontends as enabling infrastructure, and targets observable semantics rather than compiler-output identity.

M11 compatibility rule after #88

M11 must not create implementation work from a raw reference-output difference alone.

A newly discovered divergence is an M11 blocker only when evidence shows at least one of the following:

1. an observable semantic difference on the declared supported surface;
2. invalid Workshop syntax or failure of Wright's declared Workshop parse/emit round-trip contract;
3. a source identity, diagnostics, analysis, agent/source-edit, or other tooling correctness defect;
4. violation of another explicitly documented compatibility contract.

Presentation-only or upstream-canonical spelling differences remain useful N-level evidence, but they are not automatically product-critical work. Semantic compatibility is the governing criterion.

Final gate

M11 may close after all of the following are true:

1. [M11 correctness] Close supported-surface Workshop emission and round-trip divergences #87 independent verification

   • QA verifies the existing AC-17 closure boundary against the resulting code/evidence.
   • The declared matrix-surface self-roundtrip/oracle-canonical checks, corpus reports, and regression suites are green.
   • Per [M11 correctness] Close supported-surface Workshop emission and round-trip divergences #87's own decision, this is the last ad-hoc emission scan. Do not extend the issue with another spelling-only batch.

2. Architecture contract reconciled

   • [Architecture] Rebaseline Wright as a tooling-first semantic platform #88 is accepted, or at minimum the superseding ADR/compatibility interpretation needed by this gate is merged and authoritative.
   • The historical compiler-output work is not retroactively erased; its priority is simply interpreted under the new semantic-compatibility contract.

3. Evidence refreshed

   • Real-world gap inventory/support matrices reflect the current implementation, including settings support and the remaining explicitly deferred source constructs.
   • Supported, deferred, reference-limited, and inconclusive cases remain distinguishable.

4. Repository verification

   • Current main CI is green with required jobs actually executed.
   • Acceptance uses current code/tests/reports, not issue closure or implementation-agent self-report alone.

Expected gate result

accept-with-limitations is valid and expected if the declared supported surface is correct while corpus-evidenced language constructs remain explicitly deferred.

M11 does not need to make every real-world OPY project compile before closing. It must leave a trustworthy corpus-defined statement of what Wright can currently compile and analyze, which gaps block those projects, and which gaps are intentionally deferred.

Non-goals

• Broad implementation of deferred OPY syntax (\\, ++/--, dict literals, triple-quoted strings, membership operators, comprehensions, bytes literals, etc.) merely because it exists in the expanded corpus.
• Further open-ended oracle-spelling scans after [M11 correctness] Close supported-surface Workshop emission and round-trip divergences #87's declared closure boundary.
• Expanding Wright with Wright-only OPY syntax or language semantics.
• Implementing OSTW support or Workshop -> OPY/OSTW reconstruction inside M11.
• Beginning M12 implementation before M11/[Architecture] Rebaseline Wright as a tooling-first semantic platform #88 have established the new contract.

Roadmap handoff

After M11 closes:

1. [Architecture] Rebaseline Wright as a tooling-first semantic platform #88 — finish/accept the tooling-first architecture rebaseline if not already complete.
2. [M12] Tooling-first lint and Workshop stability analysis platform #89 / M12 — configurable lint and Workshop stability-analysis platform; this is the next product-value milestone.
3. [M13] Source-language interoperability and OSTW compatibility #90 / M13 — OSTW semantic compatibility and Workshop-centered source-language interoperability; keep coarse until M12 and an OSTW evidence investigation justify decomposition.

Do not mechanically continue OPY surface expansion between M11 and M12. A future OPY compatibility item should be created only when real corpus/tooling evidence shows that it blocks meaningful compilation, analysis, source tooling, or an explicit semantic compatibility claim.

[Teakowa]

PM reassessment after #83/#84/#85 (READ-ONLY)

Status of the batch (all verified):

• [M11 correctness] Preserve included-file identity for frontend diagnostics #83 — included-file diagnostic identity fixed (bee8c19, driver-only: session.rs + regression tests opy_lex_error_in_included_file_names_the_included_file / opy_include_diagnostics_resolve_through_the_registry). QA re-derived the before/after repro at f5fe337: ow1 now reports common/env.opy:7:66 (byte-true), 6v6 constants/adj_constants.opy:8:85; JSON envelope, CLI text, and LSP service paths all carry the included identity. The inventory's "span provenance defect" section is resolved. Two follow-ups already landed: test-race fix (924f636) and the -- wording correction (096891f). CI green: run 31730590203 (f5fe337) and 31731754585 (096891f, 6/6 jobs).
• [M11 contract] Align deferred .opy surface with corpus evidence #84 — support-matrix deferred surface now names \, ++/--, dict literals, """ (c57dbdb), each bullet QA-verified against observed behavior (stage, code, fixture, location); oracle claims match the pinned-9.7.10 snapshots.
• [M11 evidence] Add settings-free real-world parity candidates #85 — 0/5 settings-free candidates passed the pre-check; none acquired; no parity row (PARITY_CASES still 6). All five are settings-blocked at the native lexer; 2/5 (skirmish_elim #!obfuscate, lucioball gamemode validation) fail the pinned oracle itself — evidence for the reference-strategy investigation, not for the current surface.
• QA reassessment (a609062): no phase-1 classification changes; no class 3; the inventory's resolved span section is superseded until the next refresh.

Frequency/impact across all 14 programs: settings blocks 9/14 (dominant, always first, top-of-file); \ 2/14; dict literal 1/14 first; ++ 1/14; -- never first (3 sites); """ 1/14 (oracle fails too); #!obfuscate / OW2 gamemode 1/14 each (oracle fails too — newer-OverPy syntax).

Batch decision (one new issue, recommendation only): (a) structured unsupported-settings-block rejection diagnostic — diagnostic-only frontend change (stable code, source-located; no settings parsing, no HIR/emission change, no other-input behavior change) + synthetic regression fixture; justified by 9/14 frequency and the accidental lex-error today; (b) ride-along LSP protocol test for a lex-error in an included file under the included identity (closes the QA coverage note). Excluded with rationale: \ (2/14, explicit rejection + documented contract; implementation prohibited by non-goals), dicts/++/--/"""/#!obfuscate (non-goals), settings parsing (non-goal), oracle changes (non-goal), further candidate acquisition (pinned corpus exhausted).

Bounded reference/oracle-strategy investigation (proposal only): #85 proved the pinned 9.7.10 oracle bounds useful coverage of newer-OverPy syntax; COMPATIBILITY.md's version-range question remains open. Proposed: document the 9.7.10-to-upstream delta; build a construct x version sensitivity matrix (isolated newer-OverPy runs, nothing committed); assess evidence unlocked for the corpus; analyze machinery impact (oracle-metadata, snapshots, adapter boundary, CI); recommend keep / second-pinned-reference / re-baseline with decision criteria. Deliverables: snapshot doc + policy recommendation (ADR if adopted).

Final-gate evaluation: NOT ready — #82 stays open. Met: corpus/provenance/machinery green, no class 3, #83/#84 verified, no feature work anywhere. Pending before the gate: (1) settings-boundary structured rejection lands + verified; (2) reference-policy decision recorded from the investigation; (3) parity-baseline statement ("no new parity under current non-goals, pinned-reference bound documented") + inventory refresh. Gate acceptance criteria are scoped to those three evidence items; no new parity is required.

[Teakowa]

PM decision record — settings, oracle policy, gate (READ-ONLY, controller-executed writes)

Settings: IN for v1 (D4, typed native). The pinned oracle does not drop settings — it translates the JSONC block into a Workshop settings section before variables (verified in the 4 committed snapshots: Respawn Time Scalar: 30%, Hero Limit: Off, disabled <Mode>, enabled maps { Workshop Island }), so dropping/opaque handling is architecturally forbidden and structured-rejection alone unlocks nothing. Surface is bounded (34 keys, 4 enums, 8 modes, 5 maps, 10 heroes, 2 teams, nesting <=5, en-US only); settings are inert for HIR→WIR; contracts S/D/N. Marginal unlock: first blocker removed for 9/14; up to 4/14 fully consumable (pixelart, broken-weapons, client-to-server, inputhud — body greps clean, build verification in the batch); 3/14 partial (santa, parabola → ++; crosshair → bytes literal + concat); 2/14 oracle-blocked regardless (skirmish_elim #!obfuscate, lucioball OW2 validation — documented S divergence). Batch issue drafted and filed as #86: smallest correct set — scoped settings lexing (no global braces), v1-additive HIR node, en-US emission before variables, structured rejection for out-of-corpus surface, adapter lockstep, 3 candidate acquisitions, emission table as fixture-evidenced data with provenance; non-goals: no settings semantics, no deferred-syntax work, oracle unchanged, v1-gates unchanged.

Oracle policy: strategy 1 adopted; ADR-0007 created (status Accepted). Sensitivity matrix (Track B): 9.7.10 vs 9.7.13 delta is zero-behavior (schema data only; all 14 programs byte-identical, 8 success/6 failure); a newer reference unlocks zero additional oracle evidence; the five "inconclusive on version" flags are resolved to stable-across-versions rejections. Keep 9.7.10 sole pinned reference; pin policy: version-exact, content-pinned, change only on demonstrated behavioral need, second reference only on demonstrated divergence. No oracle migration issue — explicitly not needed. Provenance note: npm gitHead lags content by one release (content pin 889d974 vs metadata 1e268895); integrity hash keeps reproduction safe; recorded at next metadata review.

Inventory refresh (QA): new snapshot superseding m11-gap-inventory.md + the reassessment, recording the Track A surface matrix + emission finding + direction-4 decision; Track B sensitivity matrix + strategy-1 adoption + resolved version flags; updated classifications (pixelart/broken-weapons/client-to-server → pending N-level; santa/parabola → ++; crosshair → bytes/concat; skirmish_elim/lucioball → documented S divergence; others stand; zombies class 6); parity baseline: today 6 rows; expected post-batch: 3 committed + inputhud as N-level/parity candidates contingent on verification, parabola/crosshair partial.

#82 gate: NOT ready. Pending evidence for the independent final gate: (1) settings batch (#86) landed + QA-verified (AC-1..7, incl. new parity/N-level rows and S-divergence record); (2) ADR-0007 filed; (3) refreshed inventory + parity-baseline statement committed; (4) re-confirmed no class-3, all suites green. #82 stays open; no other issue changes in this decision record.

[Teakowa]

PM decision record — class-3 handling before the #82 gate (READ-ONLY)

QA's class-3 classification (string-array emission divergence, m11-inventory-post86.md at 1cd07ab) is accepted: supported-surface construct, no documented intent, complete evidence, pre-existing M8 emitter bug, not a settings regression; pixelart full-program row correctly withheld (AC-8 held).

Authorized: one bounded emitter correctness fix — wright-workshop renders string values in the oracle's Custom String("...") form wherever corpus evidence shows the oracle wraps them (known: direct Array(...) elements; shared native/adapter emission path), with regression tests (minimal repro byte-assert, pixelart full-program normalized equality, 6 gates green, 7 settings sections equal, roundtrip). Explicitly out: any deferred syntax, any normalizer masking, any v1-gates FIXTURES change.

Admin unit: filed as #87 (dedicated issue) — not inside #86. #86 closes with AC-3 recorded as contingent-not-justified; #87 carries the pixelart full-program row forward (N-level AC; PARITY_CASES row contingent on a new native==adapter HIR assertion).

Gate: #82 remains open until the fix is verified (checklist: minimal-repro parity; pixelart full-program normalized equality or precise residual divergence; gates 6/6; settings sections 256/350/511/297/136/144/476; no new class-3; oracle 20/20, adapter 22/22, differential green, six CI jobs green; inventory refresh at the fix commit). Then PM runs the final gate with a four-way decision (accept / accept-with-limitations / reject / blocked) and a comment recording verified / deferred / limitations sections. No issue is closed by this record.

[Teakowa]

PM decision — final bounded correctness batch (#87 extension) and gate path (READ-ONLY)

QA's pre-gate scan (a87b0ec, as_of 6d1417b, CI 31767535944) resolved the pixelart class-3 family (normalizedEqual True, 19,925/19,925) and surfaced three further class-3 findings on supported surface, all repro-pinned, none documented, none covered by gate fixtures: (1) trailing-if End; omission (plus the ws-parser roundtrip asymmetry — oracle-valid text rejected by 'If' requires a matching 'End'); (2) .format() constant folding (semantically equivalent, low severity); (3) non-default numeric globalvar initializers dropped (semantic — variables start at 0; the live adapter carries them, refuting the support-matrix claim). All three are authorized as one bounded correctness batch inside #87 (AC-11..AC-14): numeric initializers preserved except integer-0 (lower.rs; support-matrix corrected in-batch; new synthetic fixture with HIR parity + N-level), trailing-if emission + parser roundtrip symmetry (repro pins a-d; fixed-point test), and emitter-level constant format folding (numeric-constant args only; variable args and the HUD path untouched). Non-negotiables: no deferred syntax, no normalizer masking, no v1-gates change, no forced parity.

Gate decision criteria (unchanged): the #82 final gate runs as accept-with-limitations when the pre-gate checklist passes in full (repro parity for all three; pixelart N-level re-confirmed; gates 6/6; 7 settings sections; 12-program matrix; differential per AC-11; no new class-3; all suites + six CI jobs green; inventory refreshed) and no class-3 finding remains open. The gate comment then records verified (the complete class-3 family resolved incl. this batch; settings surface; #83/#84; pixelart N-level row; ADR-0007; parity baseline), deferred (the unchanged deferred-syntax list as post-M11 candidates), and limitations (S/D/N only; en-US; oracle 9.7.10 content-pinned; S-divergence pair; zombies class 6; first-failure evidence boundary; format-folding/trailing-if repro-pinned only; pixelart parity N-level-only). #82 stays open until the checklist passes at the batch commit.

[Teakowa]

PM decision — final remediation and gate path (READ-ONLY)

The two remaining class-3 findings from the final checklist (ea2c1a4, as_of 8182959, CI 31770753398) are accepted as classified (supported surface, repro-pinned, corpus-unexercised, zero first-failure/parity impact) and authorized as the final #87 batch: (a) emission-only placeholder canonicalization ({} → oracle's explicit {0}-style numbering, pinned from oracle repros before coding; HIR and differential untouched; HUD/constant-fold paths untouched) and (b) playervar member reads in value positions emitting the oracle's (Event Player).p with a bounded ws-parser acceptance and a roundtrip fixed-point test (SET form and method receivers untouched). Scan boundary decided: the gate's no-open-class-3 criterion is scoped to the enumerated matrix-listed emission surface — a mechanical self-roundtrip closure check (parse(emit(x)) for every producer spelling) plus oracle-canonical repro comparison; finite, enumerable from the support-matrix, and the LAST scan. Constructs outside the matrix are deferred by definition and cannot produce gate findings.

Gate criteria (unchanged, applied at the batch commit): after (a)+(b) and the closure scan verify clean (checklist: repro parity pins; pixelart normalizedEqual True; gates 6/6; 7 settings sections; PARITY_CASES 8; 12-program matrix; closure scan clean; all suites + six CI jobs green; inventory refreshed; no open class-3), the #82 final gate runs immediately as accept-with-limitations, recording verified (settings surface; #83/#84; the complete class-3 family resolved incl. placeholder canonicalization and playervar-read spelling + roundtrip; pixelart full-program N-level row; ADR-0007; parity baseline PARITY_CASES 8, v1-gates 6 unchanged; refreshed inventory), deferred (the unchanged deferred-syntax list as post-M11 candidates), and limitations (S/D/N only; en-US; oracle 9.7.10 content-pinned; S-divergence pair; zombies class 6; first-failure evidence boundary; no-open-class-3 scoped to the enumerated matrix-surface scan; pixelart parity N-level-only). #82 stays open until the checklist passes at the batch commit.