[johnml1135]

Start here: src/SIL.Machine.Morphology.HermitCrab/MorphologicalRules/AnalysisAffixProcessRule.cs -- EdgePasses and the check before the matcher call are the whole behaviour change.

Status: ready for review, based on master (c2454760), no conflicts. Landing waits for a remeasure once calibrated work-steps land on PanGloss main.

Analysis affix patterns are anchored at both ends. When the first or last pattern node is a mandatory segment constraint and the word's edge segment does not unify with it, the allomorph cannot match, so it is now rejected before the matcher runs. An optional edge always passes, and the failed-unapplication trace event is still emitted. Morpher.EdgePrefilterEnabled (internal, default on) switches it off; EdgePrefilterRejected is an internal count hook.

Measured on Sena (first 300 words of the reference list, 594 parses, one thread, same binary, flag on vs off, 4 alternating runs each): median user CPU 38.5 s -> 35.3 s (-8%), managed allocation 10.2 GB -> 8.8 GB (-14%), peak working set unchanged (153-169 MB vs 156-161 MB). The PC was loaded (load average 20-35 on 20 CPUs), so the CPU figure is noisy: per-pair change ranged from -1% to -20%. Hardware instruction counters were unavailable under WSL. 468,101 allomorph attempts skipped the matcher over those 300 words. Parse output was byte-identical in every run. Analysis steps are unchanged; the saving is cheaper attempts, not fewer. Matcher work (FST traversals plus arc checks, each prefilter unification charged one arc check; not a step count) fell by 19.55% on Sena in the earlier census.

Where to look

• Optional-edge rule: only a mandatory edge may reject, because the matcher can skip optional annotations (start anchor and Advance in TraversalMethodBase). Pinned by OptionalEdgeNode_PrefilterMatchesFullMatcher and, through a real deletion rule, ParseWord_OptionalEdgeFromUnappliedDeletion_PrefilterMatchesFullMatcher.
• Variables: an unbound alpha variable unifies with null bindings, so the check is never stricter than the matcher (AlphaVariableEdgePassesWithUnboundVariable).
• Pattern shape: only a Constraint as first or last child is checked; a leading Group gives null (GroupFirstWithOptionalLeadingConstraintDoesNotReject).
• Full parses: ParseWord_StressFixture_PrefilterMatchesFullMatcher compares 60 words with the flag on and off, sequentially and in parallel.

Not here: AnalysisRealizationalAffixProcessRule is unchanged and still runs the full matcher; the spec computes its two edge references but never reads them.

Verification: HermitCrab test project 172 passed, 0 failed; dotnet csharpier check . clean; comment-hygiene.ps1 clean. ./local_check.sh was not run (SentencePiece needs cmake).

Next: review AnalysisAffixProcessRule.cs, then approve or name an edge case to add.

---

Reading this a year from now -- start here

This is the edge-segment prefilter port tracked in #564. The census data and write-ups (result-5-edge-prefilter.md, edge-prefilter-evidence/census-net-summary.txt) are not in the tree; the figures in this description are the record. The shipped fixture early-lexicon-agglutinative under tests/SIL.Machine.Morphology.HermitCrab.Tests/Research/StressFixtures/ (generator, grammar, word list) is loaded by the full-parse test.

Evidence -- measurements and census

Same-binary A/B, flag toggled by reflection, MaxDegreeOfParallelism 1, 10 warm-up words excluded, Sena grammar and word list from the reference inputs.

Sena, 300 words	user CPU s (4 runs)	allocated MB (4 runs)
prefilter off	33.4, 39.7, 47.8, 37.3	10393, 10174, 10166, 10157
prefilter on	31.8, 39.2, 38.9, 29.9	8622, 8805, 8842, 8820

Stress fixture (60 words, 27 analyses, 3 runs): allocation 264 MB -> 123.5 MB (-53%); user CPU 2.4-2.9 s vs 2.0-3.1 s, within noise.

Earlier census, unmerged, not re-run for this edit: net matcher work fell by Indonesian 6.51%, Aweti 47.47%, Amharic 12.76%, Mbugwe 19.62%, Sena 19.55%, stress fixture 77.65%. Census trace-event counts were unchanged. Invalid or budget-aborted words were excluded from parity.

The retained size of the per-rule metadata is estimated at 108.4 KiB for the largest reference grammar (an estimate, not measured).

Surprising findings

Conformance truncate-morphotactic fixtures give different parses on repeated runs of master itself, so they cannot serve as a parity check for this change.

Closes #564

---

This change is

🤖 Generated with Claude Code

[codecov-commenter]

Codecov Report

❌ Patch coverage is 94.28571% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 74.61%. Comparing base (c245476) to head (0af3bc0).

Files with missing lines	Patch %	Lines
...rab/MorphologicalRules/AnalysisAffixProcessRule.cs	93.10%	0 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #574      +/-   ##
==========================================
+ Coverage   74.57%   74.61%   +0.04%     
==========================================
  Files         453      453              
  Lines       38342    38376      +34     
  Branches     5284     5297      +13     
==========================================
+ Hits        28593    28635      +42     
+ Misses       8583     8573      -10     
- Partials     1166     1168       +2     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[claude]

1. Verdict: approve. The round-1 gaps are closed: full parses now match with the prefilter on and off, sequentially and in parallel, including an optional edge left by unapplying a deletion rule.
2. Most important: nothing open. F1 (tests/SIL.Machine.Morphology.HermitCrab.Tests/MorphologicalRules/AnalysisEdgePrefilterTests.cs:376) was the main gap and is addressed in 614f9a7.
3. Counts: 0 Critical, 0 Important, 0 Low open; no new findings this round.
4. Ran dotnet test tests/SIL.Machine.Morphology.HermitCrab.Tests at the head: 172 passed, 0 failed. gh pr checks 574: Build on ubuntu-22.04 passed; Build on windows-latest and claude-review pending; comment hygiene passed (advisory only).
5. Not verified: that the deletion test fails when optional edges are rejected (I could not run a mutation in this environment); the census figures, which are still outside the tree; and --agent-strict comment hygiene, which I did not run.

Public API, target framework, package, machine.py parity: None verified. The new members are internal.

F1 addressed, F2 addressed, F3 accepted, F4 withdrawn, F5 addressed.

Reviewed at 0af3bc0