optic-k-rebuild

name: optic-k-rebuild description: "Rebuild the OPTIC-K voicing index (state/voicings/optick.index). Invoke when schema hash bumps, tag enrichment changes, new chord qualities are added, or the leak-test / invariant-coverage tools flag regressions. Mandatory precondition: stop any process that has the index mmap-locked — GaApi and GaMcpServer both hold the file open at runtime, and the write step WILL fail with IOException otherwise." metadata: domain: music triggers: rebuild optick, rebuild voicings, OPTIC-K rebuild, regenerate index, corpus rebuild, optick.index, v4-pp, schema hash mismatch role: operations scope: corpus-generation

OPTIC-K Index Rebuild Procedure

Operational runbook for rebuilding ga/state/voicings/optick.index. The index holds 313k guitar/bass/ukulele voicings as 112-dim OPTK v4-pp vectors. Rebuild is needed whenever:

• Schema hash changes (EmbeddingSchema.BuildCompactLayoutV4 layout string edit → CRC32 bump → old index hard-rejected by reader).
• Writer normalization semantics change (e.g. v4 global-L2 → v4-pp per-partition).
• Tag-enrichment logic changes (VoicingTagEnricher, VoicingAnalyzer, SymbolicTagRegistry).
• Chord-quality enumeration expands (new ChordPitchClasses entries that the corpus generator also uses).
• A leak-test or invariant-coverage run flags that corpus-encoded vectors drifted from the generator's intent.

Total wall-clock expectation

~140 seconds for 313k voicings on a modern dev machine. Breakdown:

• Guitar: ~136 s (667k raw → 298k unique via structural dedup → OPTIC-K embedding).
• Bass: ~2.5 s (12k raw → 7.8k unique).
• Ukulele: ~2.3 s (8.6k raw → 7.3k unique).
• Write step: seconds.

Always run in background (run_in_background: true) so the session isn't blocked. Poll the output file for progress.

⚠ Preflight — MANDATORY file-lock check

optick.index is memory-mapped by every running OptickSearchStrategy consumer. On Windows, any open mmap handle locks the file for writing. The rebuild's embedding phase succeeds normally, then the 5-second write step crashes with IOException if anything's still holding the file.

Known holders:

• GaApi.exe — the web service. Started manually or via launch profile.
• GaMcpServer.exe — spawned by Claude Code on demand via dotnet run --no-build.
• Any ad-hoc CLI or test runner that called OptickSearchStrategy and didn't dispose it.

Stop them first:

Get-Process -Name GaApi, GaMcpServer -ErrorAction SilentlyContinue | Stop-Process -Force
# verify
Get-Process -Name GaApi, GaMcpServer -ErrorAction SilentlyContinue

Or in bash:

tasklist | grep -iE "GaApi|GaMcpServer"
# then taskkill each PID
taskkill -F -PID <pid>

Do not rebuild with processes running. The embedding phase will happily burn 2+ minutes of CPU before failing at the write, producing nothing.

Backup the current index

One line, 168 MB copy, takes ~1 s:

cp "C:/Users/spare/source/repos/ga/state/voicings/optick.index" \
   "C:/Users/spare/source/repos/ga/state/voicings/optick.index.backup-$(date +%Y-%m-%d)"

If the rebuild produces a bad index (invariant regressions, search quality drops), restore from backup with cp optick.index.backup-<date> optick.index.

Build the generator

The writer lives in FretboardVoicingsCLI. Make sure its compiled DLL reflects the current writer/encoder changes before running:

dotnet build "C:/Users/spare/source/repos/ga/Demos/Music Theory/FretboardVoicingsCLI/FretboardVoicingsCLI.csproj"

Expect Build succeeded. 0 Error(s). Any MSB3027/MSB3021 errors mean a process is still holding the output DLL — recheck step 1.

Run the rebuild

From the FretboardVoicingsCLI directory (output path is resolved relative to CWD if not absolute — always pass absolute path to avoid surprises):

cd "C:/Users/spare/source/repos/ga/Demos/Music Theory/FretboardVoicingsCLI"
dotnet run --no-build -- --export-embeddings \
  --output "C:/Users/spare/source/repos/ga/state/voicings/optick.index"

CLI flags:

• --export-embeddings — mode flag (required).
• --output <abs-path> — output path; default state/voicings/optick.index relative to CWD.
• --tuning guitar|bass|ukulele — optional; omit for all three.
• --export-max <N> — optional per-instrument cap (useful for smoke tests on a sample).
• --no-dedup — disables structural prime-form dedup (produces a bigger, noisier index).

For smoke testing a pipeline change without a full 313k rebuild:

dotnet run --no-build -- --export-embeddings \
  --output "C:/tmp/optick-smoke.index" \
  --tuning guitar --export-max 2000

Verification sweep (run after every rebuild)

All four steps expected to pass. If any regresses versus the prior baseline, investigate before continuing.

1. Diagnostics snapshot

mkdir -p "C:/Users/spare/source/repos/ga/state/baseline/$(date +%Y-%m-%d)"
"C:/Users/spare/source/repos/ix/target/release/baseline-diagnostics.exe" \
  --index "C:/Users/spare/source/repos/ga/state/voicings/optick.index" \
  --out-dir "C:/Users/spare/source/repos/ga/state/baseline/$(date +%Y-%m-%d)"

Key numbers to watch (post v4-pp, pre-enrichment baseline was):

• Full-dim acc: 0.752 — classification accuracy of a forest over all 112 dims.
• STRUCTURE leak acc: 0.596 (target: drop toward 0.33 random chance).
• MORPHOLOGY acc: 0.790 — by design, instrument-correlated.
• CONTEXT/SYMBOLIC/MODAL leak acc: 0.522 / 0.614 / 0.498 (target: drop).
• Retrieval PC-set match: 92.8% (target: hold or improve).

1b. Refresh the dashboard embeddings tile (canonical snapshot)

Step 1 writes a raw report to state/baseline/<date>/ for the known-good table above — but the dashboard / ix-quality-trend read a different path: state/quality/embeddings/<UTC-date>.json. The daily embeddings-snapshot.yml cron can only ever write an amber "index absent on runner" carryforward there (the 175 MB index is gitignored, never on hosted CI). So after every local rebuild, emit the real canonical snapshot so the tile goes green instead of stale-amber:

pwsh "C:/Users/spare/source/repos/ga/Scripts/refresh-embeddings-snapshot.ps1"
git add "state/quality/embeddings/$(date -u +%Y-%m-%d).json"
git commit -m "chore(quality): real embeddings snapshot $(date -u +%Y-%m-%d)"

The cron's degraded path guards on the existing file's degraded flag, so it will not overwrite this real measurement. (Without this step the embeddings tile silently re-rots to amber the day after any rebuild — the gap that left it dark from 2026-05-16 to 2026-06-25.)

2. Invariant coverage

"C:/Users/spare/source/repos/ix/target/release/ix-invariant-produce.exe" --pretty \
  --out "C:/Users/spare/source/repos/ga/state/baseline/$(date +%Y-%m-%d)-firings.json"
"C:/Users/spare/source/repos/ix/target/release/ix-invariant-coverage.exe" \
  --catalog "C:/Users/spare/source/repos/ga/docs/methodology/invariants-catalog.md" \
  --firings "C:/Users/spare/source/repos/ga/state/baseline/$(date +%Y-%m-%d)-firings.json"

After v4-pp rebuild: #25, #28, #32 should flip from FAIL to PASS (cross-instrument STRUCTURE equality, 3-class leak test, cross-octave cosine=1.0).

3. C# integration tests

dotnet test "C:/Users/spare/source/repos/ga/Tests/Common/GA.Business.ML.Tests/GA.Business.ML.Tests.csproj" \
  --filter "FullyQualifiedName~Optick|FullyQualifiedName~MusicalQuery|FullyQualifiedName~VoicingTag"

After rebuild the previously-skipped retrieval tests in OptickIntegrationTests and OptickHardPromptBatteryTests should pass (not skip). Expect all 72 tests green.

4. Live MCP probe

If GaMcpServer is restarted (via /mcp in Claude Code), fire:

mcp__ga__ga_voicing_index_info   # liveness + new count + new schema hash
mcp__ga__ga_search_voicings { query: "Cmaj7", limit: 5 }
mcp__ga__ga_search_voicings { query: "Cmaj7 jazz", limit: 5 }

The P5 regression fingerprint flips post-enrichment: "Cmaj7 jazz" should score higher than bare "Cmaj7" (was the other way around pre-enrichment).

Rollback

If invariants regress or search quality drops:

cp "C:/Users/spare/source/repos/ga/state/voicings/optick.index.backup-<date>" \
   "C:/Users/spare/source/repos/ga/state/voicings/optick.index"

Then revert the schema-hash bump in EmbeddingSchema.BuildCompactLayoutV4 (and ix/crates/ix-optick/src/lib.rs SCHEMA_SEED) so the old index validates again.

Known-good baselines

v1.8 headline: invariant #25 flipped from 8.4% PASS (67/793) to 100% (793/793). Under v4-pp + root-boost, same-PC-set voicings across instruments had bit-different STRUCTURE slices because STRUCTURE's v[rootPitchClass] += 1.0 made the same pitch-class set differ by root. Removing the root-boost and moving root identity to a dedicated 12-dim ROOT similarity partition (weight 0.05) restored same-pitch-class-set invariance (octave / voicing / instrument — what #25 tests). NB: this is NOT full transposition (T) invariance — STRUCTURE's pitch-class chroma is T-variant (a chord vs its transposition has STRUCTURE cosine ≈0.5–0.8; measured by Tools/GaStructureInvariance); only the ICV + cardinality sub-dims are T/I-invariant.

Additional diagnostic signals in v1.8:

• ROOT partition at acc 0.345 — correctly near-random (1/3 + 1σ); diagnostic flipped its label from [LEAK] to [ok]. Root is PC-level data, genuinely cross-instrument-invariant by construction.
• β₀ range halved again (79 → 36). With STRUCTURE cleaned of root-boost, the topological geometry is even more instrument-agnostic.
• Full-dim acc DROPPED (0.867 → 0.835). Counter-intuitive but correct: cleaner STRUCTURE = less instrument signal smuggled via root-boost = less predictability of instrument from the full vector.
• Retrieval held at 100% PC-match. Adding the ROOT partition did not degrade retrieval quality.

Residual STRUCTURE leak (0.503) comes from genuine data-level cardinality differences: a 4-string ukulele cannot represent 6-note chords, so the cardinality dim legitimately differs across instruments for large PC-sets. That's an encoding-consistent truth, not a normalization bug.

Interpretation of the post-rebuild deltas:

• Retrieval 92.8% → 100.0%. Every top-K in the 50-query test now matches the query's PC set exactly. The combined effect of per-partition normalization + tag enrichment + extended quality coverage.
• CONTEXT leak completely eliminated (0.522 → 0.333, exactly random-chance). The partition no longer encodes instrument signal.
• STRUCTURE leak cut by ~13% (0.596 → 0.517). Residual signal comes from genuine cardinality differences across instruments (a 4-string ukulele can't represent the same note-count as a 6-string guitar for large chords) — not from normalization coupling, which is now clean.
• SYMBOLIC leak cut by ~14% (0.614 → 0.526) despite the enricher adding more bits. Genuinely instrument-specific tags (barre, stretch) keep some residual.
• MORPHOLOGY up (0.790 → 0.903) — by design; fretboard geometry IS instrument-distinctive.
• β₀ range halved, β₁ range collapsed 20× (178→79, 105→5). Topological structure is now far more instrument-agnostic — the per-partition norm de-couples global structure from instrument-specific details.

Invariant implications:

• #25 (cross-instrument STRUCTURE equality for same PC-set): leak test reduced substantially; a targeted same-PC-set comparison would now pass. Broader classifier still shows residual due to cardinality effects, which is a data-coverage issue, not an encoding bug.
• #28 (no partition accuracy > 1/3 + 3σ): CONTEXT now satisfies this cleanly. STRUCTURE/SYMBOLIC still slightly over threshold.
• #32 (cross-octave STRUCTURE cosine = 1.0): 100% retrieval is strong evidence this now holds in practice.
• #33 (ChordName consistency across instruments): corpus-metadata issue, unchanged by normalization fix — addressed in a separate pass.

(Add a row after each confirmed-good rebuild.)

Related

• Design doc: ix/docs/plans/2026-04-18-optic-k-v4-pp-per-partition-norm.md.
• Auto-memory: project_optic_k_v4pp.md, reference_optic_k_rebuild_procedure.md.
• Writer source: ga/Demos/Music Theory/FretboardVoicingsCLI/OptickIndexWriter.cs.
• Encoder source: ga/Common/GA.Business.ML/Search/MusicalQueryEncoder.cs.
• Reader source: ga/Common/GA.Business.ML/Search/OptickIndexReader.cs, ix/crates/ix-optick/src/lib.rs.