refactor: dedupe async/coroutine machinery (net −176 LOC)

[ivarvong]

Summary

Eight small-to-medium refactors of the async/coroutine machinery in lib/pyex/interpreter* plus two bonus cleanups. Each commit is independently reviewable and individually passes mix test + mix dialyzer.

Net change: 390 deletions, 214 insertions across 5 files. 5360 tests, 0 failures. Dialyzer clean throughout (and Unnecessary Skips: 0 after the last commit).

Commits, in landing order

1. Collapse four :cont_call_*_resume frames into one — pos/kw/star/dstar shared a frame shape and differed only in how the sent value was folded into the partial-args state. One frame + an arg_slot() discriminator + a shared Calls.incorporate_value/6 helper replace four near-identical resume handlers and four near-identical eval clauses. −126 LOC.

2. Type-separate coroutine signals from pyvalues in the yield channel — names Interpreter.coroutine_signal() for {:asyncio_sleep, _} / {:asyncio_capability_call, _, _, _} so the yield channel typechecks cleanly. Removes two .dialyzer_ignore.exs entries (:unused_fun and :pattern_match on asyncio.ex).

3. Unify :gen_awaiting_capability into :gen_awaiting_send — the two iter pool entries had identical shape and used the same step_via_send resume; the only difference (advance policy) is now encoded in the value's type. One state, one set of clauses.

4. Collapse drive_iter_to_completion and drive_after_capability — near-copy-paste drive loops differing only in advance helper and recursion target. Now a single function with a mode parameter (:buffered on entry, latches to :fresh after first cap dispatch).

5. Atomic awaitable-iter allocation — Ctx.new_awaiting_capability_iterator/4 now takes a sentinel_builder closure that receives the freshly-allocated cap_id. Removes a direct Map.put into ctx.iterators from Interpreter.call_function.

6. Name invoke + resume as Invocation.dispatch_capability/5 — small concept-naming commit. Honest about scope in the message: this is naming, not deduplication (the parallel path splits the two halves across host threads and can't use the combined helper).

7. Delete dead :builtin_kw_raw code — the tag was never constructed anywhere in lib/ or test/. Both the call_function clause and the Invocation.call_builtin_kw_raw/5 helper that pattern-matched on it were unreachable.

8. Drop 11 stale .dialyzer_ignore.exs entries — mix dialyzer --list-unused-filters reported these as no longer matching any current warning. Unnecessary Skips: 0 after.

Honest scope notes

• I claimed a 9th refactor in my original plan (always-fresh trampoline, eliminate the cap-state peek workaround). Tracing it carefully showed sleep_coroutine constructs iters that genuinely depend on buffered :gen_pending semantics; fixing that requires restructuring how sleep iters are built, which is outside dedupe scope. Flagged in the commit message for refactor Property tests with Polars oracle + interpreter optimizations #4: asyncio.sleep(0.1) wall-clocks at ~220 ms on main because the buffered sentinel re-surfaces and Process.sleep runs twice. Pre-existing on main, not fixed here.

• The unused_fun filter for lib/pyex/stdlib/asyncio.ex and pattern_match for the same file are both gone (Add crypto stdlib, default-deny network, boto3/sql capability gates #2). The remaining ignore list is now 17 entries, all of which Dialyzer confirms are still load-bearing.

Test plan

• mix format --check-formatted
• mix compile --warnings-as-errors
• mix test — 5360 tests, 0 failures
• mix dialyzer — 0 errors, 0 unnecessary skips

🤖 Generated with Claude Code