Commit 12ed40d

**Status:** KNOWN-COST, fix-class workstream queued post-terminal-goal

pure-C completion (per Alex priority order). NOT a recent regression;

stable across 11+ days, 6+ sessions, multiple unrelated commit

surfaces. Documented per supervisor 2026-04-27T15:15:20Z disposition

following pythia #195 catch on prior "pre-existing structural causal

absence" framing.

`Tools/benchmark_phoenix.py jit --only=gen_simple` reports Phoenix

(JIT_ON, `cinderjit.auto`) at ~0.70-0.76x of vanilla CPython on the

4-bench ABBA harness — i.e. JIT is ~30-37% slower than vanilla on

this single benchmark, while other benchmarks in the same run land at

positive speedup (fibonacci ~2.27x, nqueens ~1.59x, func_calls

~1.00x).

`bench_gen_simple(n_iter=8_700_000)` per-rep (`PER_BENCH_ITERS` in

`Tools/benchmark_phoenix.py`):

def bench_gen_simple(n_iter):

def gen(n):

for i in range(n):

yield i

total = 0

for _ in range(n_iter // 100): # 87,000 outer iterations

for v in gen(100): # 87,000 generator creations

total += v # 8,700,000 inner yields

return total

87,000 short-lived generators are created and exhausted per rep. The

inner `gen()` function runs ~100 yields each before the generator

object is discarded.

3-warmup + 5-measure pattern matching `_worker_jit` in

`Tools/benchmark_phoenix.py`, `n_iter=8_700_000`, `mean_ms` of 5

measurements:

| Configuration | mean_ms |

|---|---|

| vanilla CPython `-S` | 476.1 |

| Phoenix `JIT_ENABLE=0 PYTHONJITAUTO=0 -S` | 491.2 |

| Phoenix `cinderjit.auto()` `-S` | 633.2 |

`cinderjit.auto()` is the cost driver. Phoenix-with-JIT-disabled is

within 3% of vanilla (harmless CPython-branch overhead). Phoenix-JIT

adds 142ms (29% on top of JIT_OFF) on this workload.

The JIT-compiled outer loop pays per-generator-creation prologue +

epilogue cost on each of the 87,000 generators. Vanilla CPython's

interpreter has a more compact path for short-lived generator

dispatch. Specific candidates worth investigating when fix-class

workstream opens:

- JIT generator frame setup vs interpreter `_PyEvalFramePushAndInit`

- Reference-count / LOAD_DEREF cost in compiled generator frames

- `GEN_START` opcode emit weight relative to interpreter

- Inliner heuristic: does Phoenix inline `gen(100)` into the outer

loop? If not, every `gen(100)` call is full prologue.

| Date | Benchmark log | gen_simple ratio |

|---|---|---|

| 2026-04-16 | `docs/benchmarks/abba_x86_64_20260416.md` | 0.73x |

| 2026-04-16 | `docs/benchmarks/abba_arm64_20260416.md` | 0.76x |

| 2026-04-22 | `docs/benchmarks/abba_push45_20260422_040152.md` | 0.72x |

| 2026-04-22 | `docs/benchmarks/abba_pre_push46_20260422_044949.md` | 0.72-0.74x |

| 2026-04-24 | `docs/benchmarks/abba_w27c2a_2026-04-24.md` | 0.74x |

| 2026-04-25 | `docs/benchmarks/abba_24bench_push59_2026-04-25.md` | 0.70x |

| 2026-04-27 | `docs/benchmarks/abba_blme_b2_2026-04-27.md` | 0.74x |

| 2026-04-27 | `docs/benchmarks/abba_kwnames_probe_2026-04-27.md` | 0.75x |

| 2026-04-27 | `docs/benchmarks/abba_preloader_probe_2026-04-27.md` | 0.73x |

Ratio range 0.70-0.76x over 11+ days, 9 logged runs, 6+ commit

surfaces. The variance is consistent with same-session ABBA noise

floor (~3-7% per `feedback_abba_cross_session.md`); no commit in this

window introduced or removed the cost.

When per-commit ABBA reports `gen_simple` in this 0.70-0.76x range,

gate reports MUST cite this doc + use the framing

`Phoenix-stable known-cost, fix-class workstream queued

post-terminal-goal` rather than the prior dismissal language

`pre-existing, structural causal absence` (which sounded like

no-action accounting and was caught by pythia #195

2026-04-27T15:10:01Z + medic 15:15:15Z motivated-reasoning warning).

If `gen_simple` drops BELOW 0.70x or rises ABOVE 0.76x in a single

ABBA, that IS a new signal — investigate the proximate commit, do

NOT cite this doc.

1. Re-baseline `gen_simple` standalone (vanilla / Phoenix-JIT-off /

Phoenix-JIT-on) to confirm the ~0.73x ratio still holds; if not,

reframe.

2. Profile JIT-compiled `bench_gen_simple` under `perf record` or

`cProfile` (note cProfile distorts vanilla baseline by ~3.8x; use

`perf` for clean comparison).

3. Diff JIT-emitted code for `bench_gen_simple` outer loop vs

interpreter dispatch — look for per-iteration generator-frame

prologue weight.

4. Falsify each hypothesis-class candidate above; document which is

load-bearing.

5. Heavy-tier instrumentation (analogous to W-PYTORCH-CM (ii)

tp_alloc-watchpoint pattern) only after lighter probes converge or

exhaust.

- `feedback_assume_phoenix_regression.md` — Alex 2026-04-24 directive:

all bugs presumed Phoenix-introduced. Satisfied here (cost IS

Phoenix's, just not recent).

- `feedback_falsifier_convergent_negative.md` — convergent-negative

ABBA delta needs `investigated-no-quick-fix` framing, not

`pre-existing` escape. This doc IS that framing.

- `feedback_no_workarounds.md` — documenting + deferring fix is NOT a

workaround; only bail-out / deopt is the prohibited pattern.

Deferred-fix-with-workstream-anchor is transparent accounting.

- pythia #195 (D-1777300601 2026-04-27T15:10:01Z) — caught the

framing drift across 4 commits.

- medic 2026-04-27T15:15:15Z [MEDIC-WARNING] — caught generalist

follow-up framing as motivated reasoning vs pythia.