ci: run source-coverage on mods/** too -- 52 enrolled bytes the gate could never see

[andrewboudreau]

Two lines of YAML. The reason it is worth a PR is in the measurement, not the diff.

The gap

source-coverage.yml filters on src/**. But src/** is not the enrolled surface —
the enrolled surface is whatever config/**/delinks.txt names, and one of those names
is not under src/:

config/arm9/overlays/ov002/delinks.txt:1837
    mods/Player_ScaleByCharFactor.c:

52 bytes, and the only non-src/ path among the 10,973 ranges enrolled
tree-wide. A PR touching only mods/ matched neither paths: block, so the job never
ran. Not "ran and passed" — never ran. Same shape as the nearmiss/** gap closed in
#2040, and the reason it is worth chasing both: a check that does not run is rendered
identically to one that passed.

The tool itself is fine — I checked before assuming

I did not want to file a filter fix on a gate I had never seen fail, so I measured the
gate's failure branch first, at 859623fc0, in a wired worktree. Baseline: 2,067,148 B
source-built across 10,973 ranges, rombuild 106/106.

Deleting mods/Player_ScaleByCharFactor.c:

source_coverage.py -> exit 1, -52 B, blame names the file

Correct delta, correct blame. The tool was never the problem; only its trigger was.

Why I care about this particular gate

The same probe answered a question I had not seen answered anywhere: what actually
notices when dsd stops compiling a range from source and hands it back to the retail
cartridge? The answer is that it depends on how it happens, and only one gate covers
the dangerous mode:

mutation	rombuild	source_coverage
clean tree	106/106, exit 0	2,067,148 B, exit 0
function body perturbed, still enrolled	105/106, exit 1, names it	2,067,148 B, exit 0 (correctly blind)
broken body and complete removed	106/106, exit 0	2,067,084 B (−64), exit 1

The third row is the one to look at. dsd refills from the cartridge on enrollment,
never on match outcome — so un-enrolling a range produces a byte-perfect ROM. Every
byte gate green, 64 bytes silently back on the cartridge, and source_coverage is the
only thing in the tree that says so. That is exactly the row this PR was about to let
through unnoticed for mods/.

Full battery, 11 mutations, correct on all 11 — including a re-partition that hid a
4-byte hole (caught, −4), a re-partition with identical bytes (correctly silent), a
misspelled complete (−64), an un-enrolled promoted TU (−1,472), and the real
promotion commit 5c0f8b85a (no false alarm).

Stated gaps

• config_tu/ is never scanned. module_of() has explicit config_tu/ handling that
  was never wired to the scanner. 0 enrolled bytes there today, so this is latent, not
  live — filing separately rather than widening this PR.
• 92 src/ files are unenumerated by any delinks entry. They contribute 0 B, so this is
  a coverage-ceiling gap, not a refill gap.
• The metric counts enrollment, not matching, by design. Gutting a body while leaving
  complete in place gives exit 0 here — and 105/106 from rombuild, which is the
  correct division of labour between the two gates.

No compiled source touched, so there is no byte claim here and none is made. Pre-push
ran green: port_refcheck 405/405, check_src_tu 90/90, duplicate-sources 11199
stems none doubled.

[tangos-validator]

✅ PR validation — Passed

noverify: no source/build-data changes in this PR

_{Each changed src/*.c|*.cpp is compiled and its relocated bytes compared to the binary data on a private build box. Passing requires every changed file to reproduce the ROM byte-for-byte with correct relocation targets — this catches WRONG-DEST relocations and non-reproducing near-misses that ledger-scoped linkcheck skips.}