feat: motivation nudges, recognizer credit, live sol stat (0.0.53)

[okisdev]

summary

• judgment posture gains two additive motivation nudges grounded in the delegation diagnosis: the unverified-writes advisory appends a decomposition prompt when the session has zero dispatches (805 of 1395 post-demotion advisories fired in exactly that cohort), and a new once-per-session lanes reminder states the fan out default on the first prompt (new registry slug fleet-posture.session-lanes-reminder, digest 3795); judgment still never denies a main loop write
• the verification recognizer's output-summary evidence now credits vitest, jest, pytest and cargo single-line summaries in piped runs (root cause of 52 verification credits against 10880 counted writes since 07-30); a grok review of the first cut caught that go ok-line and bare "N passed" fragments would credit truncated failing tails, so those fragments are removed, cargo requires same-line "0 failed", and negative fixtures pin the truncated-failure shapes in a new dedicated suite
• the gpt-5.6-sol foreground write warning computes the real seven day timeout rate from job records (filter before cap, 16 MiB per-file gate, 5000 file budget, seven day mtime early stop, never throws) instead of quoting a hardcoded "41%" that had also survived in the codex-cli-runtime skill text; below four samples it names the 570s flight budget and the terra and luna alternatives
• the orchestration rules template encodes the diagnosis's routing policy edit candidates: verification at each coherent boundary on long inline stretches, width two is already a wave, sol on a foreground write brief needs a stated single-flight reason, and dispatching independent packages together is the unmarked case at decomposition time; rules manifest regenerated

test plan

already verified

• npm test on the merged tree -> 1129 pass, 0 fail, 1 known skip
• node --test tests/verification-command.test.mjs tests/inline-delegation-guard.test.mjs -> fail 0, including the truncated go tail and incomplete cargo negatives
• node --test tests/codex-companion.test.mjs in a full environment -> 97 pass, 0 fail, 0 skipped (the 69 process-inspection skips occur only inside peer sandboxes)
• claude plugin validate on the marketplace root and all three plugins -> 4/4 pass
• repo-wide rg "41%" -> no occurrences remain

reviewer should verify

• in a fresh judgment session on the installed plugin, the lanes reminder appears exactly once on the first prompt and not on the second
• a green vitest run piped through tail resets the write window (audit event carries evidence output-summary); a truncated failing go tail does not
• a foreground --write --model gpt-5.6-sol codex task emits the measured-rate warning with counts, or the 570s fallback below four samples