[johnml1135]

Start here: src/SIL.Machine.Morphology.HermitCrab/AnalysisAffixTemplatesRule.cs, the Walk method -- the per-template look-ahead mask on the shared trie walk. Review only the top commit (fcb65d7); the branch also carries #546 and #575.

Status: draft, stacked on #546 (shared trailing-slot trie, open) and #575 (required-slot look-ahead, open). Its base is master, so the PR diff also shows both. CI green. Closes #566.

What it does. The shared trailing-slot walk from #546 now runs #575's required-slot look-ahead at every trie node, for each admitted template below it. A template whose required slot cannot find its material is dead for that subtree while the templates sharing the node carry on; a child whose templates are all dead is skipped before its slot rule runs. On the benchmark grammars, adding the look-ahead to the shared walk cut a further 8.26% (Mbugwe), 25.97% (Sena) and 18.75% (Aweti) of work.

The unknown you arrive with: is it sound with sharing? Dead templates cannot complete (the #575 argument, applied per template), so outputs equal the shared walk without the look-ahead. The mask is one bool per template per call: no cache, no Interlocked. The same RequiredSlotLookahead Off/Shadow/On switch as #575 applies; shadow reports fired, violated and completed-under-fired counts per template.

Where to look

• Pruning one template without losing its neighbour: LookaheadPrunesOnlyTheTemplateWhoseRequiredSlotIsGone.
• No extra attempts when nothing is pruned: LookaheadKeepsEveryAttemptWhenBothTemplatesAreFeasible.
• Shadow mode on the shared walk: ShadowLookaheadOnTheSharedWalkNeverViolatesOrChangesTheOutput.
• Equality with the per-template walk on every input up to length 5: LookaheadSharedWalkMatchesPerTemplateWalkOnEveryShortInput.

Not here. The shared walk still runs only at Morpher.MaxDegreeOfParallelism == 1, as in #546; default callers get throughput from #582 (open). The look-ahead adds unification calls: 1.07% (Mbugwe) and 1.40% (Sena) more than #546 alone. Later look-ahead refinements are #576, #579 and #580.

Stack and verification. Merge order: #546, #575, then this. Ran earlier: HermitCrab test project, parity against master, #546, #575 and the combination on every benchmark word; shadow-mode violations 0. Not rerun after the rebase onto #522: the measurements below. The four truncate-morphotactic conformance failures are known.

Next: land #546 and #575, then review the top commit.

---

Reading this a year from now -- start here

This is the join of two independent optimisations: #546 removes repeated slot batches across templates, #575 prunes a partial parse as soon as a required slot's material is gone. Each alone leaves work the other removes, so the join is measured on its own (#566).

Evidence

Analysis non-template events plus synthesis steps, and unification calls including the look-ahead's own checks, master versus each variant. Measured before #522 removed the analysis memo; not rerun.

grammar	#575 alone	#546 alone	both
Mbugwe	7.07% less / 0.33% less	26.05% less / 16.09% less	32.15% less / 15.19% less
Sena	34.30% less / 1.56% less	40.09% less / 6.13% less	55.64% less / 4.81% less
Aweti	7.90% less / 4.50% less	1.61% less / 1.14% less	20.06% less / 12.18% less
Amharic	9.10% less / 0.63% MORE	0.00% equal	9.10% less / 0.63% MORE
Indonesian	equal	equal	equal

Added to #546: 8.26%, 25.97% and 18.75% less work on Mbugwe, Sena and Aweti, with 1.07% and 1.40% MORE unification calls on Mbugwe and Sena and 11.18% less on Aweti. Of the shared-slot work (about 65% of events before), 5.0% (Mbugwe), 20.2% (Sena) and 11.4% (Aweti) of events remain after both.

Shadow: 0 violations and 0 completions under a fired check on Mbugwe, Sena, Aweti, Amharic, Indonesian, conformance and the stress fixtures. Parses equal master, #546, #575 and the combination on every word.

Decisions and why

• The look-ahead runs per template, not per node: a failed template is dead for the subtree while the templates sharing the node go on.
• Same property and modes as Prune affix template partials whose required slots cannot find their material, in order and at the word edges #575, so one switch controls both.
• The dead mask is local to one call; nothing persists between calls.

🤖 Generated with Claude Code

---

This change is

[codecov-commenter]

Codecov Report

❌ Patch coverage is 89.55916% with 45 lines in your changes missing coverage. Please review.
✅ Project coverage is 74.81%. Comparing base (c245476) to head (c8df81a).
⚠️ Report is 15 commits behind head on master.

Files with missing lines	Patch %	Lines
...Morphology.HermitCrab/AnalysisAffixTemplateRule.cs	79.61%	16 Missing and 5 partials ⚠️
...hine.Morphology.HermitCrab/RequiredSlotMaterial.cs	90.00%	9 Missing and 6 partials ⚠️
...orphology.HermitCrab/AnalysisAffixTemplatesRule.cs	94.76%	7 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #577      +/-   ##
==========================================
+ Coverage   74.57%   74.81%   +0.24%     
==========================================
  Files         453      455       +2     
  Lines       38342    38752     +410     
  Branches     5284     5376      +92     
==========================================
+ Hits        28593    28993     +400     
- Misses       8583     8593      +10     
  Partials     1166     1166              

☔ 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.

[johnml1135]

Parked. This integrates #546 with the look-ahead; it will be re-cut as one commit once #546 and #575 have landed.

[claude]

1. Verdict: approve. The skip rule is sound. After an optional slot is skipped, the parent's scan of the same word, starting at that slot, already covered every slot below a pure skipped slot, with the same fired, so leaving out the re-check cannot change the result.
2. Most important: nothing new. F1 (src/SIL.Machine.Morphology.HermitCrab/AnalysisAffixTemplatesRule.cs:124) is fixed by the skipped flag, and F2 now compares against look-ahead Off.
3. Counts: Critical 0, Important 0, Low 0 new.
4. Ran: gh pr checks 577 shows both builds (ubuntu-22.04, windows-latest), the NuGet package, and comment hygiene passing; claude-review is pending. /code-review high found no correctness defect in the walk. Its other candidates are in Prune affix template partials whose required slots cannot find their material, in order and at the word edges #575 code, or are optimisations the diff gives no reason for.
5. Not verified: I did not rerun the unification counts quoted in the F1 reply, did not run the new test against the old code, and did not check the grammar measurements. The branch was rebased since bbb81a00, so the compare reports diverged. As the PR asks, I scoped this round to fcb65d7b..c8df81ac; the perf(hermitcrab): share trailing analysis template slot batches #546 and Prune affix template partials whose required slots cannot find their material, in order and at the word edges #575 code it carries belongs to those reviews.

Public API, target framework, package, and sillsdev/machine.py parity: none verified changed. IsSlotPure and WalkState are internal/private, and HermitCrab has no Python counterpart.

Findings: F1 addressed; F2 addressed.

Reviewed at c8df81a