perf(coverage): the string-scan caches cost 26ms per lookup at 121 files

[Chemaclass]

Problem

Three coverage caches store their entries in one long string and look entries up with a leading-* glob. On Bash 3.2 that scan is quadratic in the number of entries, and at real file counts the cache costs more than the work it saves.

• _BASHUNIT_COVERAGE_STATS_LOOKUP (src/coverage/stats.sh:104 get_cached_stats), report path
• _BASHUNIT_COVERAGE_TRACK_CACHE (src/coverage/engine.sh:226), capture hot path, scanned per traced line
• _BASHUNIT_COVERAGE_PATH_CACHE (src/coverage/engine.sh:242), capture hot path, scanned per traced line

Measured on Bash 3.2 arm64, macOS. report_text uses the stats cache and does nothing else expensive, yet its per-file cost grows faster than the file count:

tracked files	report_text	per file
11	0.03s	3ms
40	0.16s	4ms
80	0.85s	11ms
121	3.12s	26ms

Isolated micro-benchmark of the lookup alone, 121 entries, 8,254-byte lookup string, 605 lookups: 26.0 ms per lookup. 121 x 26ms = 3.1s, which is the whole measured report_text time. The cache is the cost.

The two capture-path caches use the same shape and are scanned once per traced line, so the same curve applies to every executed line of every tracked file.

Proposal

Replace string-scan lookup with Bash's own variable table, the associative-array emulation this repo already uses for spy state (_BASHUNIT_SPY_${variable}_TIMES_FILE, src/doubles/spy.sh):

# key -> a variable name in the reserved namespace
_bu_key="_BASHUNIT_COVERAGE_TRACK_${file//[^a-zA-Z0-9]/_}"
eval "$_bu_key=1"
decision="${!_bu_key:-}"

${!name} is Bash 2+, ${var//pattern/} is Bash 3.0+, and the paths being sanitized are short, so the substitution stays off the quadratic path that perf-fork-budget.md warns about for large strings.

For the stats cache there is a simpler fix that removes the lookup entirely: precompute_file_stats already walks get_tracked_files in order and fills the _BASHUNIT_COVERAGE_STATS_* indexed arrays. report_text should iterate those arrays instead of re-reading the tracked list and looking each file up by name.

Where to change

• src/coverage/stats.sh:66-116 the stats cache and get_cached_stats
• src/coverage/report_text.sh:39-62 iterate the precomputed arrays
• src/coverage/paths.sh:5-7 cache declarations
• src/coverage/engine.sh:215-273 record_line, both hot-path caches
• src/coverage/config.sh:90-103 cache reset in init

Acceptance criteria

• Coverage totals, per-file rows and LCOV output are byte-identical to the current implementation for the same run
• report_text cost per file stops growing with file count (measure at 11, 40, 80, 121 tracked files)
• A file whose sanitized key collides with another file's key is handled, or collisions are proven impossible for the sanitizer chosen
• Cache state is reset by bashunit::coverage::init so a second run in the same shell cannot inherit stale decisions
• Works under --parallel, where each worker builds its own caches
• Both engines produce the same numbers (BASHUNIT_COVERAGE_ENGINE=trap and xtrace)

Repo checklist (agent)

• TDD: RED then GREEN then REFACTOR. Write the failing test first.
• Bash 3.0+ only: no printf -v, no += append, no declare -A, no [[ ]], no ${var,,}, no &>>, no ${arr[-1]}. Expanding a possibly-empty array under set -u needs ${arr[@]+"${arr[@]}"}.
• Dynamic scope: everything written through eval must stay in the _BASHUNIT_* namespace, see .claude/rules/bash-style.md.
• 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).
• One issue = one PR.