Differential capture sessions¶

This index tracks original-vs-rewrite work by capture SHA256. The linked older sessions are historical investigation notes; their recordings predate the current owned formats and are not expected to load. Do not add compatibility code for them.

Start a current session only from a fresh Frida format 22 capture finalized as CDT container 2/schema 15 with its CRD replay 16 and evidence sidecars.

Session template¶

Title: Session <N> (YYYY-MM-DD)
Capture set: <native.cdt>, <run.crd>, <rng_evidence.json>, <evidence.msgpack.zst>
Capture versions: Frida 22 / CDT 2+15 / CRD 16 / evidence 2
Capture SHA256: <sha256 for every artifact>
Replay trace: <python-or-zig.cdt>
Commands: <exact health, record, diff, bisect, or focus commands>
First mismatch by channel: <channel -> tick/path>
First diagnostics: <channel -> tick/path>

Key Findings¶

<highest-signal findings only>

Landed Changes¶

<important code/tooling/doc changes only>

Outcome / Next Probe¶

<what remains, and where to probe next>

Current capture policy¶

Capture full-detail gameplay with every limit set to unlimited. Focus windows and sampled streams are not replay-grade and the finalizer rejects them.
Keep the .cdt, .crd, .rng_evidence.json, and .evidence.msgpack.zst files together as the canonical native artifact set. Log their SHA256 values and every capture knob that changes coverage.
Reject and regenerate any artifact that is not Frida 22, CDT 2/schema 15, CRD 16, and evidence 2. These files are throwaway and there is no migration path.
Run dbg health on the native trace and the replay-recorded trace before reading a diff. Stop if either selected window is not parity-ready.
Record the matching CRD with the implementation under test, then compare the two CDTs with dbg diff. Use dbg bisect/dbg focus after the first report, not as a substitute for the full-channel diff.
Log channel_first_mismatches for all affected channels. For finite float differences, record the field path, bit encodings, and f32 ULP distance.
Log channel_first_diagnostics separately. In particular, an RNG caller-attribution-only difference is useful localization evidence but not a behavioral mismatch.
If the capture SHA is unchanged, continue its existing investigation notes. Start a new session after recapturing or changing the comparison implementation.
Fixture parity is strict. Do not blanket-xfail a known mismatch; record the exact first mismatch and fix or isolate the responsible behavior.

The normal command sequence is:

uv run crimson dbg health <native.cdt>
uv run crimson dbg record <run.crd> --out <candidate.cdt> --impl <python|zig>
uv run crimson dbg health <candidate.cdt>
uv run crimson dbg diff <native.cdt> <candidate.cdt>

The raw JSONL lifecycle is strict. A capture error, tick discontinuity, or EOF with an unfinished run means recapture. EOF after run_end is complete; finalization never appends or salvages lifecycle rows.

Sessions¶

Session 18 includes continued entries in the same file.