lint: recurse W012/W013 into unobserved blocks and match arms

[Nitjsefnie]

What does this PR do?

Closes #784 and #785 — one recursion fix closes both, because W012 and W013 share a walker.

check_builtin_shadow recursed only through IF/LOOP/FOR/FUNC/TRY/PROGRAM, with AST_MATCH and AST_UNOBSERVED in its break-only group. So a shadow that warned at top level lint'd clean inside an unobserved: block or a match arm:

range is 5                    -> warning[W012]
unobserved:
    range is 5                -> silent
match 1:
    case 1:
        range is 5            -> silent

Both node types now recurse, following the shape collect_refs and collect_loop_assign_names already use for them rather than inventing one. Diagnostic codes, messages and suppression logic are untouched — this only changes where the walker looks.

Scope is deliberately these two. The sibling gaps in #780–#783 live in four other walkers (collect_assigns, check_unused_params, check_func_unreachable, check_dup_keys) and are separate changes.

Checklist

• make test passes locally — RESULTS: 3279/3279 passed, 0 failed, against a 3275/3275 baseline on main in a separate worktree. The +4 are the new tests; no pre-existing test broke.
• New builtins have signature comments and docs in docs/BUILTINS.md — n/a
• New library functions follow conventions in docs/STDLIB.md — n/a
• New examples have a comment header explaining what they demonstrate — n/a
• CHANGELOG.md updated (if user-facing change) — n/a; no behaviour change to the language, and the diagnostics themselves are unchanged. Say the word if you'd rather have an entry for the newly-firing cases.

Four regression tests in tests/test_lint.sh, in the existing check_contains fixture style, one per newly-firing case. Proven non-vacuous rather than merely passing: reverting src/lint.c and rebuilding turns all four red — Results: 114 passed, 4 failed — and restoring returns 118 passed, 0 failed.

The thing I'd most want checked in review, since a recursion widening is exactly where a linter starts crying wolf: I diffed --lint W012/W013 counts between the main binary and this branch's binary across every .eigs in lib/ tests/ examples/ tools/ bench/ web/ fuzz/. Zero differences — the fix surfaces nothing new against the repo's own sources, so it is not about to flood anyone's build.

Generated by Claude Opus 5 (brief, review), Kimi K3 (implementation)

[InauguralPhysicist]

Review — approved

Correct fix, correctly scoped. Verified against the tree, not just the diff:

• Recursion shape is right. AST_MATCH iterates case_count/body_counts[i]/bodies[i][j] and AST_UNOBSERVED uses data.block — both byte-match the shapes collect_refs already uses for these nodes, as claimed. No invented traversal.
• No missed arm. I checked eigenscript.h's match struct for a separate else/default body the recursion could skip — there isn't one; a default arm is just another pattern in patterns[], so case_count covers everything.
• Skipping match.expr and patterns[i] is correct, not an oversight. W012 fires on assignment and W013 on define, and neither can occur in a subject expression or a pattern — is is a statement. (Lambdas in expressions are already in the walker's break group at top level, so behavior stays consistent.)
• The -Werror=switch contract is preserved — the two labels moved from the break-only group to the recursing group, so the exhaustiveness gate #779 installed still covers this switch, and the comment explaining why the break group is enumerated survives.

The evidence discipline is what makes this an easy approve: red-then-green proven by reverting src/lint.c (4 red → 4 green), and the zero-diff W012/W013 sweep across every .eigs in the repo answers the crying-wolf question before I had to ask it — the widening surfaces nothing on our own sources.

One note for the sibling gaps (#780–#783): when you get to collect_assigns and friends, the same "does match have a hidden arm" and "which sub-expressions can legally contain the construct" checks apply per-walker — worth stating per PR as you did here.

CHANGELOG n/a is fine by this repo's convention (diagnostics text unchanged; only where the walker looks changed).

CI: all 16 checks green. Merging — closes #784 and #785.