Real TeX follow-up: repair scopes and explicit-definition coverage

[akcorca]

Problem and current scope

Semath is a conservative mathematical document analyzer. The practical goal is source-linked definitions and complete, safe edits in author manuscripts, plus grounded constraint checks. The broad STEM-writing program in #345 and its release plan in #354 are superseded (NOT_PLANNED), not completed.

The development corpus pins five arXiv papers (21 TeX files), preserving original source inventories, includes and author macros. See fixtures/real-documents/{sources,tasks}.json, docs/real-document-evaluation.md, and bun run corpus:collect && bun run corpus:evaluate. Original papers remain local and are not redistributed.

Annotation correction

The original 15 navigation annotations were committed before the first query (147218901a2c7d02afbfe24ec854b23f6e3d867f). Further source-context review found that primes-convolution is inside an inactive comment environment, UTF-16 offsets 8812..13517. Its former navigation-success label was invalid. The same spans now identify a separate navigation/rename refusal task.

The corrected 0.19.2 baseline is 3/14 correct navigation, 8 abstentions and 3 input rejections, plus one failed inactive-source refusal. The former 4/15 headline must not be used to compare performance. This source-label correction is not an engine improvement and was not chosen to match the output.

Current results

After #407 (Semath 0.19.3), using unchanged source archives and the corrected annotations:

Field / paper	Correct navigation	Abstained	Input rejected
Machine learning / Attention Is All You Need	1	2	0
Optimization / Adam	2	1	0
Statistics / Variational Inference	3	0	0
Theoretical physics / Origin of Gravity	0	3	0
Number theory / Small Gaps Between Primes	0	2	0
Total	6	8	0

The inactive-source task correctly refuses both definition navigation and rename, without destinations or edit proposals. There are no wrong returned destinations in this small development set. This is limited task coverage, not broad accuracy or author usability evidence. Labels are agent-authored and not independent human/blind annotations. Complete rename sets remain unscored.

Completed repairs

• Collect and pin authentic source archives; compare optimized native and committed release WASM; retain refusals/rejections in the denominator and isolate timeouts (Evaluate real TeX documents and fix concessive comparison warnings #406).
• Prevent concessive Even with k=n from becoming a global contradiction of k<n (Evaluate real TeX documents and fix concessive comparison warnings #406).
• Fix macro declaration names/replacement bodies becoming active section or environment scopes, including the gravity paper's whole-input rejection (Preserve active source and lexical scope structure wasmtex#111).
• Preserve actual section/environment containment and exclude comment/verbatim bodies from active mathematical syntax (Preserve active source and lexical scope structure wasmtex#111).
• Remove artificial per-equation-paragraph lexical scopes and make supported display-math layout transparent to lexical identity; retain theorem/section boundaries (Preserve manuscript scopes and correct source evaluation #407).
• Compare command definitions and uses using original source spelling, eliminating two incorrect unused hints in the statistics paper (Preserve manuscript scopes and correct source evaluation #407).
• Exclude label/tag metadata from mathematical symbol uses and warning evidence; the remaining t hint now points to an actual sufficient-statistic call before its following where declaration (Preserve manuscript scopes and correct source evaluation #407).
• Correct the inactive-source annotation and document the corrected historical baseline and lesson (Preserve manuscript scopes and correct source evaluation #407).

Remaining work, in order

• Support bounded coordinated declarations where source establishes identity: K/V in Attention and v_t in Adam. Keep unsupported or ambiguous evidence neutral.
• Design bounded source-preserving support for custom equation aliases. Gravity now ingests, but its \be/\ee notation is not recognized as math; all three navigation tasks abstain.
• Handle long documents within explicit analysis bounds. The number-theory M_1/M_2 declarations are beyond wasmtex's 10,000-node CST parsing budget, not missing because of theorem scope. Do not raise limits to fit the score; consider bounded region analysis with explicit truncation reporting.
• Independently label complete rename sets before claiming safe real-document edits. Review section boundaries, decorated notation and macro provenance; authorization alone is not a passed task.
• Review the three remaining use-before-definition hints (two statistics, one number theory) against original context. Add source-first shape/unit tasks only when source supplies the needed facts.

Run the same source tasks after repairs and report per-field results. Keep generated reports under .artifacts; never change expectations to match output or replace author evidence with a synthetic-law benchmark. Obtain actual author feedback before claiming time savings.

Qualification

#407 candidate 727d9b1bcad89039966e566a4e76cdd91b41b8bf passed exact x86_64 Linux release:check (reproducible WASM, check, quality, 501-document performance, package, documentation) and GitHub CI. Both clean macOS and Linux original-source runs produced identical query results and diagnostics, including native/WASM parity. WASM SHA-256: 24648ba58d30bb7a3edd4ed817d861a59f92199ac7cd7a9c657ce468f552723b.