fix(beam-search): make decode terminable, bound the grammar, add repe…

[mustfaaaa]

…tition control

DEV 2 scope: inference/beam_search.py, inference/grammar.py.

Root cause of the 0% eval: the decoder had no legal way to terminate. is_valid_prefix() required the parse to consume every token, so every continuation of a completed AST -- [EOS] included -- was rejected. The grammar mask went all -inf the moment a beam produced a correct answer, and beam_search dropped that beam from the candidate list. Only degenerate, still-open beams survived, which is why every prediction returned status="partial" with an identical "reached end of tokens" error.

Verified with an oracle stub that is certain about every gold token: before this change it still looped and never reached a solved status, so the search was broken independently of any checkpoint.

Termination

• grammar: accept [EOS] as the final token of a complete AST; add is_complete() alongside is_valid_prefix() via a shared _parse_status().
• beam_search: retire finished beams instead of discarding them; cap topk to the legal-candidate count (it was padding with -inf entries, and ties break by index, so [EOS] got selected and manufactured "finished" beams holding truncated ASTs); strip the terminator before the deserializer, which rejects any token following a closed AST.

Reachability
The grammar rejected 19.1% of real training targets: NODE:GRADIENT had no parser branch, and op-nodes could not carry the optional COEF/EXP/POINT decorators serialize_op_node emits. Because the grammar drives the decode mask those answers were unreachable, not merely mis-scored -- this is why gradient scores 0/50. Acceptance is now 100.00% of all 155,000 targets in data/splits/ (was 80.9%). Also rejects the COEF:OTHER / EXP:OTHER OOV placeholders, which parse as tokens but crash float() in the deserializer.

Repetition control (tasks 1 and 2)

• soft CTRL-style penalty on raw logits before log_softmax.
• hard guards: consecutive-run cap and a no-repeat-ngram check.
• grammar bounds: MAX_OPVARS_PER_OP, MAX_NESTING_DEPTH, MAX_SIBLINGS. Every threshold is derived from the dataset, with margin. Scoping is load-bearing: guards apply to content tokens only and never to STRUCT:CLOSE/[EOS]. 77.4% of real targets repeat some bigram but zero repeat a content-only one, so an unscoped no_repeat_ngram_size=2 would corrupt three quarters of legitimate answers.

Profiling (task 3)
scripts/profile_beam.py + docs/BEAM_SEARCH_PERFORMANCE.md. The validity mask is 4-6% of runtime at beam_size=5, not the bottleneck; the un-batched forward passes are. The grammar bounds cut the pessimistic worst case from 34.58s to 5.94s per problem. Documented bound: run_eval.py should complete in ~15 minutes and must not exceed ~30.

ONNX parity
deployment/onnx_beam_search.py had drifted -- every previous fix landed on the torch side only -- plus a numpy-only defect where np.log(softmax(x) + 1e-12) made masked tokens finite and therefore selectable. Shared policy extracted to inference/decoding.py (pure stdlib) so the paths cannot diverge again, with a parity test and a test asserting the ONNX path never imports torch (that is what keeps the Vercel bundle under the 250MB cap).

Tests: 118 -> 241 passing, 4 xfailed.

Known limitation: bounding the grammar's dimensions does not bound sequence length, so a pathological model can still fill max_len with a legal but incomplete prefix. beam_search falls back to the longest closed prefix, but that only helps when the outermost node closed. A hard guarantee needs a budget-aware closing constraint; documented, not assumed.

[vercel]

@mustfaaaa is attempting to deploy a commit to the seno-quantum-coder's projects Team on Vercel.

A member of the Team first needs to authorize it.