feat: execution arbiter default-on for debug/review (v3.12.0) - #13

The change

-if [[ "${SYNOD_EXEC_ARBITER:-0}" == "1" && ( "$MODE" == "debug" || "$MODE" == "review" ) \
+if [[ "${SYNOD_EXEC_ARBITER:-1}" == "1" && ( "$MODE" == "debug" || "$MODE" == "review" ) \

One character. Everything else is documentation catching up.

Why the flag was redundant

The gate already required all three of:

• MODE is debug or review
• a TARGET_PATH was given
• the Phase 0.5 probe collected ≥ 1 test (exec_arbiter.py:57-62)

Those conditions select exactly the situations where execution can settle
something. The opt-in flag on top was suppressing a signal the pipeline had
already qualified — so the default path kept settling code disputes by argument
even when a runnable suite was sitting right there.

Execution-grounded selection is how SWE-bench SOTA picks among candidates
(CWM, arXiv:2510.02387), and it is the
strongest mechanical signal available — the direction 3.8–3.10 moved every other
decision signal.

== "1", not != "0"

This step executes code, so an unrecognised value must fail toward not
running. Verified against bash rather than assumed:

This also matches the repo's dominant idiom (SYNOD_ANONYMIZE:-1 == 1,
SYNOD_V2_AUTO_CLASSIFY:-1 == 1).

What this means in practice

In debug/review mode against a repo with a collectable pytest suite, Synod now
runs that suite. pytest imports conftest.py and every test module during
collection, so set SYNOD_EXEC_ARBITER=0 for targets whose suite has real side
effects (live services, shared databases, outbound mail).

Bounds are unchanged: pytest -x -q, 120s hard timeout, timeout treated as
UNSETTLED rather than failing, and the arbiter never blocks the pipeline.

Known limitation — documented, not pre-solved

The arbiter runs the target suite with no green baseline, so a pre-existing
failing test is reported as machine-verified evidence (-x stops at the first
failure, which on a red repo is likely unrelated to the dispute). Recorded under
Unreleased → Planned to be fixed if false signals actually show up in use —
building it speculatively, along with scoping the run to the disputed tests, was
over-engineering the owner correctly pushed back on.

Verification

• Gate truth table exercised in bash (table above)
• 975 passed / 14 skipped; tests/test_version_consistency.py green
• All three JSON manifests parse; version 3.12.0 in all four declarations
• ruff check + ruff format --check clean
• Docs updated in lockstep: SKILL.md flag table, Phase 2 module heading + guard note, both READMEs' research-basis rows, marketplace copy that still said "opt-in"

🤖 Generated with Claude Code