lint: recurse W002 into unobserved blocks and match arms

[Nitjsefnie]

What does this PR do?

Fixes #781. check_unused_params recursed only through AST_IF/FUNC/FOR/PROGRAM and kept both AST_MATCH and AST_UNOBSERVED in its break-only group, so a define nested in an unobserved: block or a match arm was never visited and the W002 emission was never reached.

Both arms were missing here, so this adds both — unlike #790, where the walker already recursed AST_UNOBSERVED and only AST_MATCH was needed. The recursion idiom is #787's for the same node types, verbatim.

Changes

• src/lint.c — two recursion arms added to check_unused_params's switch; both labels removed from the break-only group. No new traversal helper, no change to the emission site, no other walker touched.
• tests/test_lint.sh — three tests, placed mid-file next to the existing W002 coverage rather than appended at the tail, since lint: W003 (unreachable code) never fires for a function defined inside unobserved: or a match arm #782 and lint: W010 (duplicate dict key) never fires inside unobserved: or a match arm #783 are adding tests to this same file. Two cover the fixed behaviour; the third pins the top-level path that already worked, so a future change to this switch can't silently break it.

Testing

• bash tests/test_lint.sh run directly for its exit code: 123 passed, 0 failed, 123 total, exit 0.
• Full suite tests/run_all_tests.sh: RESULTS: 3315/3315 passed, 0 failed.
• Both reproducers from the issue go from silent to warning; the control still warns exactly once.
• The tests bite. With src/lint.c reverted to upstream and the new tests kept, both behaviour tests fail (121 passed, 2 failed). The third test passes against unfixed code, which is what a regression pin should do.
• Neither arm is dead. Deleting AST_MATCH alone silences the match reproducer while unobserved: still warns; deleting AST_UNOBSERVED alone is the exact mirror.
• Verified against the merge result, not just the branch: rebased onto current main (1750b48, which now includes lint: recurse W001 collect_assigns into match arms #790), builds clean, tests/test_lint.sh exit 0.

One caveat stated plainly, because it is easy to over-read. Linting all 397 .eigs files under lib/ + examples/ + tests/ with the unfixed and fixed binaries produces byte-identical output (112,051 bytes, 973 warnings, per-file exit codes included). That is a no-regression result and nothing more: an indentation-aware scan finds zero defines nested inside unobserved: or match anywhere in that corpus, so even a no-op arm would have produced the same bytes. The two new tests are the only evidence the fix actually fires. fuzz/corpus/ was deliberately outside that scope.

Checklist

• make test passes locally — full suite via tests/run_all_tests.sh: 3315/3315 passed, 0 failed; lint tests run directly for their exit code, 123 passed, 0 failed, exit 0
• New builtins have signature comments and docs in docs/BUILTINS.md — n/a, no new builtins
• New library functions follow conventions in docs/STDLIB.md — n/a, no new library functions
• New examples have a comment header explaining what they demonstrate — n/a, no new examples
• CHANGELOG.md updated (if user-facing change) — not done, following the two merged precedents for this fix family: lint: recurse W012/W013 into unobserved blocks and match arms #787 and lint: recurse W001 collect_assigns into match arms #790 each changed only src/lint.c and tests/test_lint.sh. Happy to add an entry if you'd rather these carried one.

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