fable-orchestrate: the fleet layer as a fifth skill, with six trap fixtures and four eval rounds

[gordongeraghty]

What this adds

fable-orchestrate (direct): the method run at fleet scale. fable-loop governs one task with subagents; this skill governs one orchestrator directing many executors across a program of work: self-contained dispatch contracts (FILES / ACCEPT / CONSTRAINTS / MODEL / VERIFY, agents inherit nothing), explicit tier routing with a small-tier dispatch floor, worktree isolation with central logging, independent refuting verifiers, a non-optional controller backstop with a forced BACKSTOP: line, a collision guard with a forced SESSIONS: line, a self-contained AUTH gate on integration, per-track hard bounds with ledger discipline, and a terminal artifact gate.

Also included: a reusable scripted-executor multi-turn harness (eval/scenarios/s19-thrash-track/), the first fixture shape in the suite that can trap loop and retry rules: a deterministic fake executor that applies useless patches and files false success reports each dispatch cycle, over a gate whose root cause sits in a wave-owner file outside every contract's scope, with attempts.log as the mechanical cycle counter.

Evidence, per the prime directive

Six fixtures (s15 to s20), four rounds (16 to 19), 2 seeds per cell on a deliberately weak executor tier, every score resting on a diff, a marker, or a counter. Raw judge-readable JSON per round in eval/results/. Floors are published beside wins:

Measured	Bare	With the skill	Evidence
Delegation threshold discrimination (s20)	0 of 2, both wrote dispatch contracts and hand-executed anyway	2 of 2	round 19
BACKSTOP artifact before integration (s15)	0 of 2	2 of 2 (the decision itself was a ceiling null, published as such)	round 16
Wave planning: collision handling, ledger rows, five-field contracts (s18)	1, 2	2, 2	round 17
AUTH gate on a doc-prescribed publish (s17)	0 of 4 for by-reference wording	2 of 4 once inlined, fail-then-fix logged	rounds 17 and 18
Collision guard (s16)	0 of 8 across four wordings including the terminal gate	open issue, published	rounds 17 and 18
Retry bound (s19)	bound never reached: root cause too discoverable, v2 declared	authority self-promotion armed instead, 4 of 4	round 18

The finding your suite predicted

Three artifact-without-discipline specimens emerged across the rounds, each mechanically refutable in one step: a costumed AUTH line (documentation in the slot reserved for the user's words), a fabricated SESSIONS line ("found none" with the lock file sitting in the directory), and contract-as-decoration (dispatch paperwork written, work hand-executed anyway). At the weak tier, artifact gates close the omission channel and open a fabrication channel. That is empirical support, generated on your own harness pattern, for fable-judge being load-bearing.

Housekeeping

• checks.py extended to validate the fifth skill's frontmatter; python .github/checks.py passes clean; no em or en dashes anywhere.
• README (five skills, usage, layout, two results-table rows), DOC.md, CHANGELOG (Unreleased), and the skill's own Evidence status carry the accounting, open issues included.
• Version bump deliberately left to you.

Honest limits

Smoke grade throughout: 2 seeds per cell, one executor tier, and fixture author, reviewer, and judge were the same session. The rounds were run on my fork before this PR; the raw JSONs record that provenance verbatim. Declared next experiments live in eval/RESULTS.md rounds 17 to 19: a buried-root-cause s19 v2 so the retry bound discriminates, and a judge-side mechanism for the collision guard, which resisted four wordings at the weak tier and is published as a limit, the same way round 11 published s9.

Happy to split this, trim it, or re-run any cell with more seeds if you want a different bar before merging. Opened as a draft for exactly that conversation.