infer an empty collection literal's type from the first value stored into it

[kacy]

summary

mut tasks := [] left the binding untyped: an empty literal has no element type, so it checked as an error and every later use of the name was silently untyped. the visible damage was awaited task results. a task pulled out of such a list lost its result type, so (await t).unwrap_or(...) compiled to a bare method call the runtime cannot resolve, and r.ok emitted a field read with no offset, which the ir contract check rejects at build. the payload kind did not matter — int, string, and list payloads all broke in this shape — while every payload kind worked with a direct t := spawn f() or an annotated list.

the checker now retypes the binding from the first value stored into it (push for a list, insert for a map, add for a set) and records the type on the literal node, so the emitter builds a tagged container that retains its elements. storing none or another empty literal decides nothing; the binding stays open until a typed store arrives.

the grpc bench client carried a workaround for exactly this: workers pushed latencies into shared list handles because task results with list payloads were avoided. each worker now returns its own latency list through its task result and the caller merges after the joins.

what was tested

• new regression tests/cases/test_task_result_payloads.pith: int, string, bytes, list, map, struct, bool, and float payloads through the previously broken inferred-list shape, in both spellings (unwrap_or and .is_ok/.ok), with ! propagation and catch on top; verified on green default, one green worker, and PITH_GREEN=0
• new leak case tests/leaks/leak_task_result_payload churns both extraction arms (ok payload handed out, err released with the fallback built in place) and is flat across 200k/800k rounds; the regression also joins the curated memcheck list and runs valgrind-clean with PITH_STRUCT_FREELIST=0
• make test, PITH_GREEN=0 make test, verify-green-corpus and verify-osthread-corpus (283 passed each), make memcheck, make leak-check, make run-examples, make docsite-check, make bootstrap-verify; the bootstrap seed was regenerated and rebuilds a working driver from scratch
• examples/grpc_chat.pith and examples/grpc_reflect.pith run to completion; the bench client ran against the local tls server at conc=1 and conc=8 and prints the same row format as before

notes

a pre-existing leak surfaced while checking ownership, unrelated to this change: rebinding one name to results with different payload kinds in a single function releases later bindings with the first binding's kind, so r := await ta (an Int!) followed by r := await tb (a Bytes!) leaks the bytes payload. it reproduces with direct spawns on main. the regression test uses distinct names per section to stay out of its way; it deserves its own fix.