test: close the sprint's coverage gaps, and put its ownership fixes under memcheck

[kacy]

A coverage audit of #804-#812 found the sprint's ownership fixes were invisible to every gate that could catch their real failure mode. This closes the gaps; no compiler or runtime changes.

the memcheck hole

None of the sprint's goldens ran under valgrind, and #804/#805/#809/#810 are ownership fixes whose regression shape is an over-release or overread — which prints the right answer, stays flat in the leak gate, and passes the golden diff. Seven cases join MEMCHECK_CASES (all verified clean under the gate's exact invocation): test_generic_enum_match, test_closure_many_captures, test_closure_capture_boundaries, test_optional_match_ownership, test_optional_payload_eq, test_tail_match_defer, test_argument_literal_ownership.

new and extended tests

• test_tail_match_defer (new): fix: give a match on an rc-payload optional exact ownership #809's core mechanism — a tail match returning through the full return contract — had no test with a defer in it; a revert to the bare ret would have skipped every defer silently. Covers the expression arm, the none arm, and a borrowed-field return.
• test_closure_capture_boundaries (new): exactly 16 captures (no spill), 17 (first spill), and a 39-capture closure growing the extension from a lower high-water mark, with strings sitting on the 15/16 boundary.
• test_generic_enum_match (extended): a two-parameter base with a mixed generic+concrete payload list (Both(A, B, Int)), matched through a function boundary — per-slot kind substitution had only single-param coverage.
• tests/invalid/generic_enum_nonexhaustive (new): generic-enum matches are under E204; this is the tree's first E204 fixture at all.
• test_generic_impl_name_collision (new): std.iter's MapIter and a local generic MapIter registering in one compile — the exact collision E263's first cut had, previously caught only transitively by bootstrap-verify. Also pins the legal inherent+interface pair in both orders. (Deliberately does not drive std's MapIter: a same-named generic instance still misbinds cross-module at runtime — the pre-existing E209/feat: expand concurrency foundations #163 family.)
• tests/invalid/duplicate_generic_impl_method (new): a genuine same-kind duplicate on a generic base within one module still errors.
• test_optional_widen_eq (extended): the plain-value-on-the-left != corner, and a plain argument widening into a fully generic T? parameter.
• leak_closure_spill_capture (new leak-gate case): 20 captures, four heap strings in the spill region, closure created and dropped per round — flat (-12kb over 600k extra rounds). This is the case that gates "a spilled slot carries its release tag".
• leak_argument_literal (extended): the non-empty nested literal poke([["seed"]]) joins the round (flat, 64kb noise). The stored-tuple shape (keep((i, i+1)) into a struct field) was measured, leaks its box per round as docs/ownership.md already documents ("real tuple literals still strand their box"), and is deliberately kept OUT of the gate with a comment saying why — its correctness side stays covered by test_argument_literal_ownership.

what was tested

• All seven newly wired memcheck cases pass valgrind with the gate's exact flags (PITH_STRUCT_FREELIST=0, --error-exitcode=99).
• Both touched leak cases measured with the gate's own sequence at 200k/800k rounds: leak_closure_spill_capture -12kb, leak_argument_literal +64kb.
• make check-invalid-only 56 passed; make run-regressions-only 374 passed (the three new goldens included).
• fmt run on every new/changed pith file.