Repository navigation
v0.3.0
Highlights
- Add the VALORANT 13.08 payload transform and its native-vector evidence
from PR #32. - Fix all fourteen correctness findings recorded in
docs/CODE_REVIEW.md,
including input protection, actor attribution, RPC identity, bounded field
reads, explicit terminators, zero-width leaves and asset-reader validation. - Fix Windows Parquet input and test-fixture ownership without cleanup retries,
sleeps, skipped tests or reduced suite counts.
Compatibility notes
This is a minor release because downstream adapter output semantics change:
- RPC grouping now includes nullable
object_net_guid. Distinct subobjects
no longer merge into one invocation. rpc_receivedretains object identity. Shot events preserve null for actor
rows instead of substituting the actor GUID. Consumers that need an effective
target must choose their actor fallback explicitly.- Present zero-width array leaves are retained and can produce additional child
rows on affected inputs. - Plant attribution at an unordered custody transition or terminal close remains
unresolved and makes the spike-carrier command fail rather than guess a player.
Repeated calls on the same object in one packet remain a separately counted
adapter limitation. Existing unnamed-property and property-key-collision losses
remain visible; conversion is not generally claimed to be loss-free.
Correctness and diagnostics
- Refuse spike output paths that resolve to any source input; retain safe
replacement of earlier analysis outputs and the atomic hard-link contract. - Preserve packet and physical row ordering for equal-time ownership writes.
- Retire the displaced actor after an empty failed channel reopen, without
inventing a close event. - Bound effect tags and values to their declared payload windows.
- Require consumed RepLayout and movement-array terminators while retaining
emitted prefixes, valid sparse updates and raw failure evidence. - Report components and bits outside a sized movement window separately in
main/checkpoint summaries and manifests. Quality gates reject missing or
nonzero observations without inventing measurements in historical baselines. - Reject short or oversized intermediate IoStore blocks, unbounded method-name
tables and repeated directory file identities. - Measure documentation test counts for the executing platform.
- Close Python analysis inputs explicitly, including filtered movement scans,
FastArray auxiliary tallies, shared single-file reads and streaming evidence
readers (PR #36).
Verification scope
Before the version-only increment, the final fixes passed all nine hosted CI
checks, including Windows Rust 1.86, Rust stable, Python 3.12/3.13 on Windows
and Ubuntu, and optimized Windows packaging.
The three pinned public 12.10, 12.11 and 13.00 fixtures passed main/checkpoint
audits and 582 independent typed/raw comparisons with no width failure or value
mismatch. All 39 public Parquet files were byte-identical before and after the
correctness changes. Local suites passed 737 non-Windows Rust and 1074 Python
tests; the Windows suite contains 738 Rust tests.
The ownership follow-up also passed all nine PR checks and the required manual
Windows candidate CI. Real-file regressions retain inputs and assert closure
before cleanup; disabling closure makes those checks fail.
The tagged commit passed every required check. The release workflow tested
the optimized Windows CLI and movement decoder, ran the extracted executable,
and audited the pinned 12.10 replay with that packaged binary.
The published ZIP and SHA-256 sidecar are verified again before upload;
build-info.json binds the tag, source commit, target and executable hash.
The private 1,018-replay corpus and the two private 13.08 replays were not rerun
for this release. The historical 13.08 contributor audit reports one strictly
clean replay out of two; the other has two case-only checkpoint GUID-path
disagreements. No new native-reader capture or installed-game asset extraction
is claimed. Public fixture success does not measure every malformed-edge
frequency or exercise positive plant attribution; deterministic synthetic CLI
regressions cover that attribution boundary.
What's Changed
- Add VALORANT 13.08 by @SunkenInTime in #32
- fix: resolve CI file locks and recorded correctness findings by @yakisoba0728 in #34
- release: prepare v0.3.0 by @yakisoba0728 in #35
- fix(tools): own all single-file Parquet inputs by @yakisoba0728 in #36
New Contributors
- @SunkenInTime made their first contribution in #32
Full Changelog: v0.2.5...v0.3.0