bpo-28331: fix "CPython implementation detail:" label is removed when content is translated. by methane · Pull Request #195 · python/cpython

Stackless can now pickle coroutine_wrapper objects.

…r coroutines (python#195)

Silently ignore attempts to close a running generator, coroutine or asynchronous
generator. This avoids spurious error messages, if such an object is deallocated
as part of a paused, restorable tasklet.
(cherry picked from commit 905d0ef)

Investigation artifact per supervisor 15:14:01Z (in-tree evidence
prevents next-session/future readers from re-discovering). Closes the
'pre-existing structural causal absence' dismissal pattern that BLME
b2 / kwnames / preloader_ ABBA reports applied to gen_simple's 0.73x
ratio (pythia python#195 catch).

  docs/benchmarks/gen_simple_investigation_2026-04-27.md

3-config focused timing at ABBA-calibrated n_iter=8,700,000:
  Vanilla CPython 3.12          505 ms  1.00x
  Phoenix interpreter (no auto) 472 ms  0.94x  (FASTER than vanilla)
  Phoenix + cinderjit.auto      670 ms  1.33x  (SLOWER than vanilla,
                                                1.42x SLOWER than
                                                Phoenix's interpreter)

The 0.73x ratio is the Phoenix JIT generator-compilation cost itself
— not pre-existing baseline noise. Phoenix's interpreter beats vanilla
on this workload by 6%; only the JIT path regresses. The 4 observer
commits (BLME b2, kwnames, preloader_, push-59 baseline) don't touch
generator codegen so they couldn't have introduced the regression —
but the regression IS Phoenix-introduced, just upstream of those
probes.

Per feedback_assume_phoenix_regression.md (Alex 2026-04-24): every
regression presumed Phoenix-introduced; per
feedback_falsifier_convergent_negative.md: convergent-negative delta
needs investigated-no-quick-fix framing, not pre-existing escape.

Mechanism (hypothesis): Phoenix cinderjit.auto JIT-compiles the
inner gen() closure once the call-count crosses the 1000-call
auto-compile threshold. The JIT'd generator runs slower per-yield
than the interpreter on this workload — likely candidates: per-yield
codegen overhead in Phoenix's JIT generator path, or auto-compile
threshold + counting overhead amortized across too few iterations.

Not yet root-caused at the Phoenix-source layer. Next-step options
(deferred to next-session fix-class workstream per supervisor
15:14:01Z):
  (a) cinderx_dev oracle bisect (Tier-1 falsifier per Alex's policy
      — recommended first step per theologian 15:13:57Z)
  (b) Profile + perf-record the JIT'd gen() execution
  (c) Park as 'investigated-no-quick-fix' if (a)+(b) show
      architectural cost with no localizable hotspot

ABBA-report convention going forward (per supervisor 15:14:01Z
cascade): gen_simple is no longer 'pre-existing'; new framing is
'Phoenix JIT generator-compilation regression, root-cause class
identified (this artifact), bisect queued for next-session'. Until
the bisect lands, this artifact is the workstream-tag the gatekeeper
>5% BLOCK rule (15:12:00Z) accepts as the evidence-class root-cause
citation distinguishing the regression from a generic pre-existing
dismissal.

Doc-only commit. No JIT/build/codegen surface; trivial-push
exemption.

Stable 0.70-0.76x vs vanilla CPython on bench_gen_simple across 11+
days, 6+ commit surfaces, 9 logged ABBA runs since 2026-04-16. Not a
recent regression.

Mechanism candidate: 87,000 short-lived generators per rep × per-
generator-creation JIT-prologue/epilogue cost. Confirmed
cinderjit.auto IS the cost driver (Phoenix JIT_OFF 491ms vs JIT_ON
633ms vs vanilla 476ms; testkeeper /tmp/gen_simple_match_harness.py
2026-04-27T15:14:16Z).

Doc supersedes the prior dismissal framing 'pre-existing structural
causal absence' (caught by pythia python#195 + medic motivated-reasoning
warning). Replacement framing: 'Phoenix-stable known-cost, fix-class
workstream queued post-terminal-goal'. Anti-pattern guard + threshold
signal preserved (drops below 0.70x or rises above 0.76x = NEW
signal, investigate).

Status: PARKED behind pure-C JIT terminal-goal completion per Alex
priority order. Resumption gate enumerates 5-step procedure when
fix-class work opens.

Per supervisor 2026-04-27T15:15:20Z disposition (a) following
testkeeper investigation 15:14:16Z + theologian patch-shape APPROVE
15:17:13Z.