lint: W001 (unused variable) never fires inside a match arm

[Nitjsefnie]

Describe the bug

W001 (unused variable) is never reported for an assignment that exists only inside a match arm. collect_assigns recurses into AST_BLOCK and AST_UNOBSERVED (src/lint.c:344-348) but lists AST_MATCH in its break-only group (src/lint.c:385), so the name is never recorded and the emission loop at src/lint.c:3061-3077 never sees it.

Unlike the sibling gaps, this one is match-only — unobserved: bodies are walked and do report correctly.

Surfaced while doing #779, which made these switches enumerable; it neither caused nor fixed this. Line numbers resolved against main @ 23edbc25.

To reproduce

Control — reported correctly:

w001_top is 5

Also reported correctly (unobserved: is walked):

unobserved:
    w001_inunobs is 5

Silent — the bug:

match 1:
    case 1:
        w001_inmatch is 5
    case _:
        w001_inmatch2 is 6

Expected behavior

warning[W001]: unused variable 'w001_inmatch' (and 'w001_inmatch2'), as the top-level and unobserved: cases already produce.

Actual behavior

$ eigenscript --lint w001_control.eigs
w001_control.eigs:1: warning[W001]: unused variable 'w001_top'
$ eigenscript --lint w001_unobs.eigs
w001_unobs.eigs:2: warning[W001]: unused variable 'w001_inunobs'
$ eigenscript --lint w001_match.eigs
w001_match.eigs: no issues found

The program runs (exit 0), so this is live code the linter is silent on, not a parse artifact.

Environment

• OS: Debian 13 (x86_64)
• GCC version: 14
• EigenScript version: 0.34.0 (main @ 23edbc25)

[InauguralPhysicist]

Confirmed — verified the walker inspection for all four of #780–#783 (the three define-level walkers break on both AST_MATCH and AST_UNOBSERVED; collect_assigns handles unobserved: but not match arms, matching your narrower claim here), and spot-checked this one behaviorally: deadvar is 42 in a match arm lints clean while the same line at top level fires W001.

All four are open to a PR — same contract as #787: recursion following the collect_refs shape, red-then-green regression tests, and the zero-diff W-count sweep over the repo's own sources. One PR per walker or one PR for the quartet is your call; the per-walker "which sub-expressions can legally contain the construct" check from the #787 review applies to each.

Labeling the four bug / good first issue.