perf(coverage): emit the whole LCOV report in one awk invocation

[Chemaclass]

Problem

After #1085 and #1088 the report phase is report_lcov, and it is the only
place left that still forks per file and loops per line in Bash.

Per tracked file it runs three awk forks (lines, functions, and the branch
extractor once it moves) plus two Bash loops: the per-function span scan that
decides FNDA, and the per-arm scan that decides BRDA.

Measured on Bash 3.2 arm64, 128 tracked src/*.sh files:

	128 files
report_lcov total	2785 ms
of which compute_branch_hits	1479 ms
of which extract_functions	535 ms
of which awk_lcov_lines	440 ms
bare awk fork x128 (the floor)	~350 ms

The remaining cost is almost entirely fork overhead and Bash glue: the awk
work itself is a rounding error, as #1088 showed when the same shape took the
stats phase from 2585 ms to 153 ms.

Proposal

One awk invocation for the whole report, driven by the same manifest #1088
introduced. Every record a file needs — SF, FN/FNDA/FNF/FNH,
BRDA/BRF/BRH, DA/LF/LH, end_of_record — comes out of a single
pass that already has the source lines and the propagated hit counts in hand.

This needs the function scanner and the branch extractor in function form so
both the per-file API and the batch pass call the same code, not a copy.

Verification

Byte-identical LCOV against a fixed corpus, with hits, plus the existing unit
tests. The branch extractor has already been ported and diffed against the
Bash implementation over all 463 shell files in the repo with no mismatch.

[Chemaclass]

Prototype done, measured on the same 128 tracked src/*.sh files with hits:

	128 files
reference, per file (today)	3133 ms
one awk invocation	262 ms

Output is byte-identical: 15,357 LCOV records, every SF/FN/FNDA/
FNF/FNH/BRDA/BRF/BRH/DA/LF/LH/end_of_record line in the same
order.

One trap worth recording for the port: the function scanner keeps its
single-quote character in BEGIN { SQ = sprintf("%c", 39) }, because a literal
' cannot appear inside the single-quoted shell string that holds the awk
source. Composing the scanner into another program without carrying that
BEGIN over leaves SQ empty, so quotes stop being tracked and the scanner
silently regresses to the pre-#1087 behaviour — it does not error. Whatever
shape the shared chunk takes, SQ has to travel with bu_scan.

Scope for the change:

• the scanner helpers (bu_scan + SQ) and the branch helpers move into the
  shared awk chunk, so the per-file API and the batch pass run the same code
• report_lcov writes the perf(coverage): compute every file's line stats in one awk pass, not one Bash loop per file #1088 manifest and makes one call
• the per-file path stays as the fallback, as in perf(coverage): compute every file's line stats in one awk pass, not one Bash loop per file #1088
• a differential test pins the awk branch extractor against the Bash
  extract_branches over every shell file in the repo, the way perf(coverage): move line classification and LCOV emission into awk #1059 pinned
  the classifier (already run by hand: 463 files, 0 mismatches)

[Chemaclass]

Done in #1091 (merged).