Repository navigation
HermitCrab analysis: combined stack cut work 14-64% on four real grammars, but R2 is closed; re-measure without it (unmeasured) #611
Description
Claude-drafted.
The combined stack of #546, R1, R3, R2 and R5 cut work by 14.08% to 64.31% on four real grammars, but R2 is now closed, so the stack must be re-measured without it. Against master, less work (analysis events plus synthesis steps): Mbugwe 33.26%, Sena 64.31%, Aweti 27.58%, Amharic 14.08%, Indonesian equal, conformance 2.01%. The synthetic stress grammar bantu-tam: 89.64%. Parses equal master on every completed word. These numbers exist only in the retired PR #490 ledger. Measured before R2's witness-reuse fix.
R1 is the required-slot look-ahead (run on #546's shared walk via #577), R3 its ordered and anchored form (#576), R2 the root-substring prune (#572), R5 the edge-segment prefilter (#574).
Status: R2 (PR #572) closed 2026-10-29. Counting every unit of work as equal it added 7.7% to Sena's total work (closing comment on #572).
What we know:
- R2 adds 4.74 points of Sena work on top of perf(hermitcrab): share trailing analysis template slot batches #546 with R3 (59.57% to 64.31% less). On Mbugwe R2 is off, so the stack equals perf(hermitcrab): share trailing analysis template slot batches #546 with R3.
- R2 standalone on conformance was 15.3% more work (events, each probe counted as one event; ledger row 40).
Not known: the stack's numbers without R2.
Proposed change (C#): rerun the stack on the current branches in work steps, record the numbers in this issue, and list the merge order. Use PR #578 for the stress grammars.
Related: PRs #546, #577 (draft), #574, #578 (open); #576 (merged); #572 (closed); #562, #567, #566 (open). PanGloss: sillsdev/PanGloss#112 (the pointer issue for this one, open); sillsdev/PanGloss#66 (open) ports #546 and #566; sillsdev/PanGloss#23 (open) ports #572.
Next: rerun the stack without R2 in work steps and post the table here.
Evidence -- the full stack table
One census binary, one harness, one word set, every arm against master; census mode, 150 s per word; words aborted or erroring in any arm are dropped from all arms. "Less" is against master, per grammar.
| grammar | work | unification | matcher |
|---|---|---|---|
| Mbugwe (40 words) | 33.26% | 20.96% | 38.31% |
| Sena (51) | 64.31% | 15.43% | 22.45% |
| Aweti (99) | 27.58% | 45.64% | 55.71% |
| Amharic (24) | 14.08% | 10.42% | 14.96% |
| Indonesian (70) | equal | 5.49% | 6.51% |
| conformance (1,452 runs) | 2.01% | 3.18% | 8.69% |
| stress bantu-tam (60) | 89.64% | 64.03% | 81.38% |
| stress early-lexicon (60) | 67.72% | 83.86% | 87.06% |
Parity: 0 differences in per-word parse signatures; 0 shadow violations on the full stack. The stack needs one test-only commit (11 toy-grammar tests set RootSubstringPrune off).
Additivity (work, points of master): Mbugwe 33.26 against a sum of 34.23 (near-additive); Sena 64.31 against 103.30 (overlap: #546, R3 and R2 remove the same dead branches); Aweti 27.58 against 13.12 (super-additive).
R2's cost -- why it is closed
- Standalone (with the witness-reuse fix): Sena 9.05% more unification, conformance 1.00% more, bantu-tam 37.48% more. Events: Sena 17.29% fewer, conformance 0.22% fewer.
- Before the fix, net of probes: Sena 15.3% less work, conformance 15.3% more (7,237 probes buy 92 events).
- Stacked, the fix removes about 90% of R2's unification overhead, but R2 still costs 1.03 to 3.71 points of unification work over no R2 (conformance, Sena, bantu-tam).
- The ledger's landing decision: Prune analysis rule outputs that contain no root when no later rule can restore stem material #572 lands only on top of perf(hermitcrab): run the required-slot look-ahead on the shared trailing-slot template walk #577 and Skip impossible analysis affix matches using edge segments #574. The alternative, dropping Prune analysis rule outputs that contain no root when no later rule can restore stem material #572, gives up Sena's 4.74 points.
Sources -- pinned at 2eba2f5 (closed PR #490)
- Analysis round, "Combined stack and R2's cost" (from line 364)
- Ledger row 47 (line 179) and row 40 (line 172)
- Closing comment on PR Prune analysis rule outputs that contain no root when no later rule can restore stem material #572 (source of the 7.7% Sena figure; not in the retired record).
Activity
[-]HermitCrab analysis: record the combined stack (#546 + R1 + R3 + R2 + R5) and its merge order[/-][+]HermitCrab analysis: combined stack cut work 14-64% on four real grammars, but R2 is closed; re-measure without it (unmeasured)[/+]