fix(audio/oracle): stop global alignment locking onto a capture's silent tail (nightly #418) #140

Merged

robocub merged 2 commits from fix/oracle-align-overlap-floor into master 2026-10-01 16:05:32 +00:00

Conversation 0 Commits 2 Files changed 8 +212 -33

robocub commented 2026-10-01 14:05:07 +00:00

Member

Copy link

Why. Nightly #418 (2026-09-29) failed audio_d2m_unenc with mean NCC 0.009 although the audio crossed intact: replaying the artifact WAVs through the real oracle shows cycles 2–7 at ≥ 0.98 once aligned correctly. align() refined each coarse candidate over the first second of the source only; a slip inside cycle 0 zeroed the true lag's score and a lag overlapping ~120 ms of tail dither (score 0.047) won, so every per-cycle window pointed at silence.

What changed (crates/audio/src/oracle.rs):

• coarse candidates must overlap at least half the shorter signal (was 100 ms) — tail lags of a long capture are excluded outright;
• refinement correlates over 3 s instead of 1 s;
• candidates are ranked by raw NCC over their whole overlap (capped at 30 s), not by the refinement-window score.

Evidence.

Capture	before	after
run 418 (failed night)	delay 15.87 s, mean NCC 0.009, 0/8	delay 3.794 s, mean NCC 0.860, 5/8 → PASS
run 416 (passing night)	delay 3.716 s, 0.818, 6/8	unchanged

New regression test alignment_survives_a_damaged_first_second (long lead-in, damaged first cycle, 4 s dither tail) fails on the old code and passes now. Workspace clippy -D warnings clean; 378 tests pass on forgejo-runner. Also adds crates/audio/examples/oracle_replay.rs, the offline replay used for the diagnosis.

Harness-only change; no bridge runtime code touched. A live e2e dispatch of the four audio scenarios on this branch is linked below.

🤖 Generated with Claude Code

**Why.** Nightly #418 (2026-09-29) failed `audio_d2m_unenc` with mean NCC 0.009 although the audio crossed intact: replaying the artifact WAVs through the real oracle shows cycles 2–7 at ≥ 0.98 once aligned correctly. `align()` refined each coarse candidate over the first second of the source only; a slip inside cycle 0 zeroed the true lag's score and a lag overlapping ~120 ms of tail dither (score 0.047) won, so every per-cycle window pointed at silence. **What changed** (`crates/audio/src/oracle.rs`): - coarse candidates must overlap at least half the shorter signal (was 100 ms) — tail lags of a long capture are excluded outright; - refinement correlates over 3 s instead of 1 s; - candidates are ranked by raw NCC over their whole overlap (capped at 30 s), not by the refinement-window score. **Evidence.** | Capture | before | after | |---|---|---| | run 418 (failed night) | delay 15.87 s, mean NCC 0.009, 0/8 | delay 3.794 s, mean NCC 0.860, 5/8 → PASS | | run 416 (passing night) | delay 3.716 s, 0.818, 6/8 | unchanged | New regression test `alignment_survives_a_damaged_first_second` (long lead-in, damaged first cycle, 4 s dither tail) fails on the old code and passes now. Workspace clippy `-D warnings` clean; 378 tests pass on forgejo-runner. Also adds `crates/audio/examples/oracle_replay.rs`, the offline replay used for the diagnosis. Harness-only change; no bridge runtime code touched. A live e2e dispatch of the four audio scenarios on this branch is linked below. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

robocub added 1 commit 2026-10-01 14:05:07 +00:00

fix(audio/oracle): stop global alignment locking onto a capture's silent tail ... df5b22ad38

Nightly #418 (2026-09-29) failed audio_d2m_unenc with mean NCC 0.009 while
the audio had crossed intact (cycles 2-7 >= 0.98 once aligned). align()
refined each coarse candidate over the first second of the source only, so
a slip inside cycle 0 zeroed the true lag's score and a lag overlapping
~120 ms of tail dither (score 0.047) won; every per-cycle window then
pointed at silence.

- coarse candidates must overlap at least half the shorter signal (was
  100 ms), which removes tail lags of a long capture outright;
- refinement correlates over 3 s instead of 1 s;
- candidates are ranked by raw NCC over their whole overlap (capped at
  30 s), not by the refinement-window score.

Regression test: long lead-in + damaged first cycle (35 ms slip, and a
phase-inverted cycle) + 4 s of tail dither; fails on the old code (delay
lost), passes now. Replaying the failed night's WAVs through the fixed
oracle: delay 3.794 s, mean NCC 0.860, 5/8 -> PASS; the passing night is
unchanged (0.818, 6/8).

Also adds crates/audio/examples/oracle_replay.rs, the offline replay used
for that diagnosis.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

robocub force-pushed fix/oracle-align-overlap-floor from df5b22ad38 to 2fc038b0a9 2026-10-01 14:08:23 +00:00 Compare

robocub added 1 commit 2026-10-01 14:20:41 +00:00

ci: allow clippy::double_must_use on the two async_trait items (Rust 1.99 clippy) ... 9425fd3110

Rust 1.99.0 (2026-09-28) reached the CI toolchain today; its clippy denies
double_must_use under -D warnings on the #[must_use] that the async_trait
macro attaches to its generated boxed-future methods, breaking every
workspace clippy run. The attribute is the macro's, so the lint is allowed
on exactly those two items.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

robocub force-pushed fix/oracle-align-overlap-floor from 9425fd3110 to 8ff2bf3f69 2026-10-01 14:22:05 +00:00 Compare

robocub merged commit f4ec2503eb into master 2026-10-01 16:05:32 +00:00

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels   

bug

Something is not working

  

duplicate

This issue or pull request already exists

  

enhancement

New feature

  

help wanted

Need some help

  

invalid

Something is wrong

  

question

More information is needed

  

wontfix

This won't be fixed

No labels

bug

duplicate

enhancement

help wanted

invalid

question

wontfix

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

nether/nether-voicebridge!140

No description provided.