perf(coverage): move line classification and LCOV emission into awk

[Chemaclass]

Problem

report_lcov is a Bash loop over every line of every tracked source file, and it is the largest single cost in a coverage run.

Measured on Bash 3.2 arm64, macOS, report phase only, synthetic hits:

tracked files	report_lcov	share of default report
11	0.91s	72%
40	2.66s	69%
121	9.93s	61%

At 121 files the default report phase (precompute plus text plus LCOV) is 16.4s, of which 9.93s is LCOV emission. src/ is 18,751 lines, and the emitter walks all of them in Bash while calling is_executable_line per line.

.claude/rules/perf-fork-budget.md already records the rule this violates: a while read loop lost to a single awk fork by 5x on 600 lines, and file scans stay awk for that reason. LCOV emission is an 18,000-line file scan written as a Bash loop.

Proposal

Move line classification and LCOV emission into one awk program per run.

• Port the is_executable_line rules (src/coverage/lines.sh:21-115) to awk. The rules are pure text tests, no shell state.
• Feed awk the tracked-file list plus the aggregated hit counts, emit the whole LCOV file in one pass.
• Keep the Bash implementation as the reference for the differential test, not as a second code path shipped to users.

The classifier's documented quirks must be reproduced exactly, or coverage numbers move silently. .claude/rules/perf-fork-budget.md prescribes the method used for #1005: run both implementations over every git ls-files '*.sh' line, compare per line, and mutation-test the harness so a differential that cannot fail is not mistaken for a passing one.

Two known traps from #1005, both of which change numbers without erroring:

• inside a POSIX bracket expression a backslash is a literal member of the set, so [^\)] also excludes \
• [\{\}] matches a lone \, so a bare line continuation counts as a brace-only line

Note the ordering with the classifier bug fix: if the case-pattern rule is corrected first, port the corrected rules and diff against the corrected Bash version.

Where to change

• src/coverage/report_lcov.sh the emitter
• src/coverage/lines.sh:21-115 the rules being ported, kept as the reference implementation
• tests/unit/coverage/executable_test.sh the rule tests, which must pass against both implementations
• new differential test covering every tracked *.sh file in the repo

Acceptance criteria

• LCOV output is byte-identical to the current implementation for the same run, across the whole repo as input
• A differential test compares Bash and awk classification for every line of every git ls-files '*.sh' file and fails on any disagreement
• The differential harness is mutation-tested: deliberately breaking one rule in one implementation must fail the test
• report_lcov at 121 tracked files is measurably faster, before and after numbers in the PR description
• Fork count does not grow with source-line count, the budget shape tests/acceptance/bashunit_coverage_forks_test.sh already uses
• Works on the awk implementations CI covers, including BusyBox awk on Alpine and macOS awk
• $AWK is used, not a bare awk, so the pinned binary rule holds (src/config/env.sh)
• Both engines produce the same numbers (trap and xtrace)

Repo checklist (agent)

• TDD: RED then GREEN then REFACTOR. Write the failing test first.
• Bash 3.0+ only in the shell parts: no printf -v, no += append, no declare -A, no [[ ]], no ${var,,}, no &>>, no ${arr[-1]}.
• Keep the awk program in a single-quoted variable like _BASHUNIT_COVERAGE_XTRACE_EMIT_AWK (src/coverage/engine.sh:132) so the build's line-length and quoting rules hold.
• Gates: make sa, make lint, ./bashunit tests/, ./bashunit --parallel tests/. Never run shfmt -w.
• CHANGELOG.md: one line under ## Unreleased (performance entry with the measured factor).
• Consider an ADR: this changes which language owns the classification rules.
• One issue = one PR.

[Chemaclass]

Done in #1080 (merged).