chore: tracking — tree-sitter/ast-grep structural-parsing offload sweep

[laurigates]

Umbrella for the tree-sitter / ast-grep offload sweep. Origin: a portfolio review of which plugin skills and repo scripts still do structural parsing with regex/grep — exactly the places with documented false-positive histories (#1701/#1721/#1722/#1900 in the bash hook; #1744/#1492 in the markdown lints). The unifying principle is .claude/rules/offload-to-deterministic-substrate.md: match code/markdown by structure with a real parser, keep the agent for judgment.

Items

Suggested implementation order (best effort-to-payoff first): #2010 → #2009 → #2013, then #2011/#2012 opportunistically; #2008 is explicitly gated.

Tier 1 — repo hooks & lint scripts (accuracy)

• feat(hooks-plugin): ast-grep bash-parse fast path for bash-antipatterns.sh (gated on next false-positive) #2008 — bash-antipatterns.sh ast-grep bash-parse fast path — gated: do not implement until the next false-positive regression, then implement instead of another regex patch
• feat(scripts): shared tree-sitter-markdown extraction helper for the five markdown-structure lint scripts #2009 — shared tree-sitter-markdown extraction helper; migrate the five fence/table/blockquote-parsing lint scripts

Tier 2 — code-quality-plugin skills (operational simplification)

• feat(code-quality-plugin): ship code-antipatterns + code-hidden-failures catalogs as ast-grep rule projects #2010 — ship code-antipatterns + code-hidden-failures catalogs as executable ast-grep rule projects (ast-grep scan -c … --json); agents keep only judgment work
• feat(code-quality-plugin): replace code-complexity manual AST analysis with lizard offload #2011 — code-complexity: replace "manual AST analysis" branches (JS/TS, Go) with lizard
• feat(code-quality-plugin): offload dry-consolidation duplicate discovery to jscpd + ast-grep #2012 — dry-consolidation: offload duplicate discovery to jscpd (+ ast-grep for exact-shape confirmation)
• feat(code-quality-plugin): route code-targeted bulk sweeps in bulk-sweep-classify through ast-grep #2013 — bulk-sweep-classify: route code-targeted sweeps through ast-grep, keeping the four-category discipline for prose/docs

Explicitly out of scope

• migration-patterns-plugin (eslint-to-biome, flake8-to-ruff, …) — config migrations; the target tools ship their own migrators (biome migrate eslint). No AST work to offload.
• Raw tree-sitter library bindings anywhere except the markdown lint helper (feat(scripts): shared tree-sitter-markdown extraction helper for the five markdown-structure lint scripts #2009) — ast-grep (single binary, aqua-installable) covers everything else.
• code-quality-preflight-cue.sh export regex (code-quality-preflight-cue over-triggers on small standalone shell scripts #1766) — already fixed per-filetype; a parser there is overkill for a nudge hook.

Cross-cutting conventions for all items

• Every skill change lands with its semantic regression guard (.claude/rules/regression-testing.md)
• New tool dependencies follow the install priority (mise aqua/pipx, uv tool install); runtime hooks must fail open when the tool is absent, CI-only scripts may hard-depend
• Docs land in the same commit as code (blueprint-plugin:blueprint-docs-currency)

[laurigates]

Wave 1 landed as 5 open PRs (all green, mergeable)

A parallel worktree workflow implemented the five actionable items. Each PR carries its semantic regression guard and lands docs in the same commit.

Item	PR	Notes
#2010	#2019	code-antipatterns (13 rules) + code-hidden-failures (9 errors-track rules) shipped as ast-grep rule projects with ast-grep test fixtures. Shell track stays grep-based by design. CI (test-skill-scripts) now installs ast-grep so the fixture gate actually runs.
#2009	#2020	scripts/lib/extract-md-elements.py (uv PEP723 tree-sitter-markdown helper); migrated check-context-command-execution.sh (#1744) + check-version-pin-coverage.sh (#1492), behavior-preserving. The three fixed-string denylist lints deferred with rationale in the PR.
#2011	#2018	code-complexity JS/TS + Go manual-AST branches replaced with lizard.
#2012	#2017	dry-consolidation discovery offloaded to jscpd + ast-grep, Grep fallback retained.
#2013	#2016	bulk-sweep-classify routes code-targeted sweeps through ast-grep; four-category prose discipline retained.

#2008 remains intentionally un-started — the parking spot for the bash-antipatterns.sh structural fast path, gated on the next false-positive regression. Not implemented preemptively.

Checkboxes will be ticked as each PR merges (child issues auto-close via Closes #N).