fix: reclaim the tuple-literal argument box, and tag nested argument literals

[kacy]

The last two measured leaks from docs/ownership.md's argument-position family, both probe-confirmed live on main before fixing.

the tuple-literal argument box (~64 B/call)

f((a, b)) stranded one box per call. The kind alone can never justify a release — an optional and a result lower to the same tuple shell, so a tuple-TYPED register might be a wrapper whose count the extraction paths still hand on — but a syntactic (a, b) in the source is unambiguous: the box is one only this call ever saw, and a storing callee takes its own count. The owned-argument release (ir_release_owned_string_operand) now fires for exactly that node shape; any other tuple-typed register keeps the old bounded leak rather than risking a double free. (The first cut of this gated on ir_rc_kind returning empty for "tuple" — it doesn't, it returns "tuple", which would have released every wrapper in argument position. Caught before it ever built; the landed gate is on the AST node kind inside the caller.)

the nested argument literal (f([[]]), ~30 B/round)

A literal WITH elements types itself, but an empty literal nested inside leaves the whole expression error-typed — and an error-typed argument skips the coercion path the type-mismatch check would have run, so the inner list stayed untagged and never owned the strings pushed into it. propagate_empty_argument_collection_type now runs collection_literal_coerces for that shape: the element walk qualifies the literal honestly and records the declared type through every nested literal. A genuinely wrong element still fails the coercion and reports as before.

docs/ownership.md updates both entries; with these merged, every leak that file documented with a measured per-round figure in the argument family is closed.

what was tested

• probes at 300k rounds, before → after (peak RSS): f((i, i+1)) 20.8 MB → 2.2 MB; f([[]]) with a string push per round 11.2 MB → 2.2 MB; outputs identical before and after
• new tests/cases/test_argument_literal_ownership.pith golden: the released tuple, a callee that stores the tuple into a struct and reads it back after the caller-side release (no dangle), the tagged nested empty literal, and a nested literal with elements
• new tests/leaks/leak_argument_literal gate case, measured with the gate's own sequence: flat at 200k vs 800k rounds (2656 vs 2696 kb peak)
• make bootstrap-verify: bootstrap ir fixed point verified (includes self regressions + std tests)
• 8 consecutive make sitegen-check runs clean — the golden-output gate that catches over-releases
• make check-invalid-only: 51 passed, 0 failed; fmt + lint clean; bootstrap seed regenerated