feat(coverage): report real per-arm execution counts in BRDA #1061

Problem

Branch coverage records whether an arm was taken, never how often. compute_branch_hits (src/coverage/branches.sh:223) emits taken_count as 0 or 1, documented as an MVP limit, and report_lcov writes that straight into BRDA (src/coverage/report_lcov.sh:64).

LCOV consumers read the BRDA fourth field as an execution count. Reporting 1 for an arm taken 5,000 times and 1 for an arm taken once loses the distinction that makes branch data useful: which paths are exercised in earnest and which are grazed by a single test.

The per-line data needed already exists. _BASHUNIT_COVERAGE_HITS_BY_LINE holds real execution counts, which is what LCOV DA records emit today.

Proposal

Report the arm's execution count instead of a taken flag.

• _arm_taken (src/coverage/branches.sh:204) returns on the first hit line. Instead, take the count of the arm's first executable line, which is the closest analogue to how other tools count an arm entry.
• Decide the rule explicitly and document it: first executable line's count, or the maximum across the arm's lines. First-line count is the honest one for an arm, since later lines can be skipped by an early return.
• BRDA then carries a real count, and BRH still counts arms with a non-zero count, so existing summaries stay valid.

Low priority next to the correctness issues in this audit. It improves data quality, it does not correct a wrong number.

Where to change

• src/coverage/branches.sh:196-251 _arm_taken, compute_branch_hits
• src/coverage/report_lcov.sh:60-69 BRDA emission
• src/coverage/html_file.sh if the HTML report shows branch data

Acceptance criteria

• BRDA reports the arm's execution count, not 0 or 1
• An arm inside a loop that ran 10 times reports 10
• An arm never taken still reports 0
• BRF and BRH totals are unchanged for the same run
• The chosen counting rule is documented in the function comment
• genhtml accepts the output, tested on a real LCOV file
• 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: no printf -v, no += append, no declare -A, no [[ ]], no ${var,,}, no &>>, no ${arr[-1]}.
• The _branch_* helpers mutate their caller's locals through dynamic scoping on purpose and must stay in one file, see ADR-010 and the comment at src/coverage/branches.sh:196-202.
• Gates: make sa, make lint, ./bashunit tests/, ./bashunit --parallel tests/. Never run shfmt -w.
• CHANGELOG.md: one line under ## Unreleased.
• One issue = one PR.