perf: fix top 5 high-performance-go.md violations from a 3-agent audit

[jeduden]

Summary

Three parallel subagents audited origin/main against
docs/development/high-performance-go.md,
each covering a distinct slice of the guide (allocations/strings-bytes,
struct-layout/skip-work, concurrency/anti-pattern-table). This is the fifth
round of this recurring audit (after #762, #764, #765, #767), so most
low-hanging fruit was already gone; the findings below are what survived
verification against the current code.

Ranked by severity/impact, the top 5 real, safely-fixable issues found:

1. markdownflavor (MDS034) fixGitHubAlerts / buildAlertSkipMaps —
   both converted every scanned line to a string (one via an unsized,
   growing []string accumulator, the other via a per-continuation-line
   strings.TrimLeft(string(...))). Both now stay in []byte, and
   fixGitHubAlerts pre-sizes its accumulator. 176 → 70 allocs on an
   addPrefix-heavy 50-alert fixture (measured against an origin/main
   worktree of the same fixture, not the original, less representative
   fixture the first commit shipped with — see the review-round commits for
   that correction).
2. cmd/mdsmith --build-verify's snapshotOutputs — read every
   declared build output fully into memory via os.ReadFile to diff two
   verify-pass runs byte-for-byte. Declared outputs are user-defined
   artifacts of unbounded size; this held two full copies of a large output
   per verify pass. Streams each output through sha256 instead. Behavior
   change, intentional: the old bytes.Equal(nil, []byte{}) comparison
   treated a missing output and a present-but-empty one as the same "no
   content" case; the new hash comparison treats them as different, which
   is real non-determinism --build-verify exists to catch. Documented on
   outputHash and pinned by TestOutputsEqual_MissingVsEmpty_ReturnsFalse.
3. maxsectionlength (MDS036) collectParagraphs — ran its own
   ast.Walk (duplicating the walk astutil.CollectSectionParagraphs
   already performs and memoizes per File for MDS023/MDS024) and called
   ExtractPlainText + CountWords per paragraph for a value that never
   needed the materialized text. Now delegates to the shared memoized walk
   and uses mdtext.CountWordsInNode. 36 → 1 alloc on a 30-paragraph
   fixture (the single remaining alloc is the presized result slice; the
   walk itself is a cache hit after the first call within a File).
4. slidevstructure/recipesafety pure-membership sets —
   builtinLayouts, knownFMKeys, and bodyTemplateReserved were
   map[string]bool read only via truth-tests. Converted to
   map[string]struct{} per the guide's "map[K]struct{} for sets" pattern.

A sixth finding — MDS003/MDS005 each running a redundant full ast.Walk
instead of joining the engine's shared KindScopedChecker dispatch — was
the highest-severity hit (both rules are default-enabled and run on every
file with headings), but converting them safely needs a per-file
state-reset hook the NodeChecker contract doesn't currently provide (a
rule instance is reused sequentially across files by its worker, and
CheckNode has no "start of file" signal). Shipping that conversion
without solving the reset problem risks a state-leak correctness bug
across files, so it's tracked in
plan/2608020650_kindscoped-heading-rules.md
instead of included here. Four more rules share the same double-walk shape
and blocker (blockquotewhitespace, codeblockstyle, samefileanchor,
requiredstructure) and are noted in that plan as follow-ups.

Each fix follows red/green TDD: a failing allocation/memory-bound test
first, then the minimal fix to make it pass. Three independent xhigh-effort
adversarial review rounds ran against this diff after the initial 5 fixes;
each found real issues (a missing -race test skip, an under-exercised
test fixture, a misleading/unnecessary package var, a duplicated AST walk,
a tautological test, an undocumented nil-vs-empty difference, a leftover
string() conversion, two inaccurate code comments, and a stale
short-circuit-order regression) — all fixed, commit-by-commit, in the
review-round commits below.

Test plan

• go build ./...
• go test ./... (all packages green)
• go test -race on every touched package
• go vet ./...
• go tool -modfile=tools/go.mod golangci-lint run (0 issues)
• mdsmith check . (568 files, 0 failures)
• New/updated tests per fix: markdownflavor alloc + correctness
  tests (including a lazy-continuation indent edge case) plus a
  race_on_test.go/race_off_test.go pair, builddiag bounded-memory
  test + outputHash-typed unit tests including the missing-vs-empty
  distinction, maxsectionlength alloc test,
  slidevstructure/recipesafety zero-byte-set reflection tests

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.71%. Comparing base (2ab4b29) to head (13354cc).

Additional details and impacted files

Components	Coverage Δ
Go	98.70% <100.00%> (+<0.01%)	⬆️
TypeScript	99.54% <ø> (ø)

☔ 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.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[Copilot]

Pull request overview

This PR applies several verified performance fixes across mdsmith’s lint rules and CLI build verification path, aligned with the project’s docs/development/high-performance-go.md guidance (alloc avoidance, skipping unnecessary work, bounded memory, and set representation). It also records a follow-up plan for the higher-risk “KindScopedChecker reset” design needed to safely eliminate redundant AST walks in stateful heading rules.

Changes:

• Reduce allocations in markdownflavor GitHub-alert fixing by staying in []byte and pre-sizing accumulators.
• Make --build-verify output comparison bounded-memory by hashing outputs instead of reading full files into RAM, with regression tests (including missing-vs-empty behavior).
• Reuse memoized AST paragraph collection in maxsectionlength and convert pure-membership “set maps” to map[string]struct{} with guard tests.

Reviewed changes

Copilot reviewed 17 out of 17 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
plan/2608020650_kindscoped-heading-rules.md	Adds a plan to design a safe per-file reset mechanism before converting stateful heading rules to KindScopedChecker.
PLAN.md	Registers the new plan in the repository plan index.
internal/rules/slidevstructure/setsize_test.go	Adds tests pinning set maps to map[string]struct{} value type.
internal/rules/slidevstructure/rule.go	Converts builtinLayouts/knownFMKeys from map[string]bool to map[string]struct{} and updates membership checks.
internal/rules/recipesafety/setsize_test.go	Adds a test pinning bodyTemplateReserved to map[string]struct{}.
internal/rules/recipesafety/rule.go	Converts bodyTemplateReserved to map[string]struct{} and updates membership checks.
internal/rules/maxsectionlength/rule.go	Reuses astutil.CollectSectionParagraphs memo and counts words directly from AST nodes.
internal/rules/maxsectionlength/paragraphs_alloc_test.go	Adds an allocation regression test to prevent reintroducing text extraction / redundant walks.
internal/rules/markdownflavor/rule.go	Reworks GitHub-alert fix path to avoid per-line string() conversions and to pre-size output.
internal/rules/markdownflavor/rule_test.go	Adds a direct-return test for the nil-vs-empty behavior of fixGitHubAlerts.
internal/rules/markdownflavor/race_on_test.go	Introduces raceEnabled build-tag constant for race gating.
internal/rules/markdownflavor/race_off_test.go	Introduces raceEnabled build-tag constant for non-race builds.
internal/rules/markdownflavor/fix_test.go	Adds coverage for indented lazy-continuation rewriting behavior.
internal/rules/markdownflavor/alloc_test.go	Adds/updates allocation budget test for fixGitHubAlerts.
cmd/mdsmith/buildpass_diag_test.go	Updates tests to validate output hashing and the missing-vs-empty distinction.
cmd/mdsmith/builddiag.go	Switches build-verify output snapshotting from os.ReadFile to streamed sha256 hashing with stricter equality semantics.
cmd/mdsmith/builddiag_memcap_test.go	Adds a bounded-memory regression test for hashing large build outputs.

[Copilot]

Pull request overview

Copilot reviewed 17 out of 17 changed files in this pull request and generated no new comments.

Suppressed comments (1)

cmd/mdsmith/builddiag_memcap_test.go:51

• This memory-cap test uses runtime.MemStats.TotalAlloc (a global counter) while this package also has several t.Parallel() tests (e.g. lsp_unit_test.go), so unrelated concurrent allocations can be counted in grew and make this assertion flaky. Consider loosening the threshold so it still catches the old os.ReadFile behavior (≈8 MiB alloc for an 8 MiB file) while tolerating background/parallel-test noise.

	// The whole point of the fix: total heap growth must stay well under
	// the output's size, proving the content was streamed through a
	// hash rather than read whole into a []byte.
	grew := after.TotalAlloc - before.TotalAlloc
	if grew > size/4 {
		t.Fatalf("snapshotOutputs on an %d-byte output allocated %d bytes; "+
			"want well under the file size (no full-file buffering)", size, grew)
	}