trajctl dogfood — the witness report

Issue: #2541 (spine step 8 of epic #2533). File the gaps, do not fix them.

This is the witness step of the trajectory-control spine: the whole declare → score → curve → (steer) spin exercised on one real long-horizon session, with real ledger rows and at least one regime-gate decision. It is deliberately honest about what could not be exercised — that miss list is the point of a dogfood.

The session under test

The real long-horizon session is the trajectory-control epic build-out itself (#2533): steps 1–7 of the spine landed across many turns as a sequence of real commits on trunk. That is a genuine multi-turn objective with witnessed progress, so it is the right thing to score. The objective was declared with the seven spine steps as plan phases, each phase bound to the real commit that landed it:

phase	spine step	landing commit
phase-1	objective score ledger spine (#2534)	41e62c0b97dc
phase-2	declare/close/list objective CLI (#2535)	0fdd64fa95b1
phase-3	scorer registry + W3 witnessed-commit progress (#2536)	07f7f2f738ab
phase-4	W2 activity-divergence (stall) scorer (#2537)	8f23d7f3bc18
phase-5	curve fold + closed signal vocabulary (#2538)	4c7e852d2436
phase-6	score-at-turn-end runner + PreCompact twin (#2539)	f649954e4d84
phase-7	re-anchor nudge — regime-gated steering (#2540)	(none — #2540 is OPEN)

How it was run

The turn ends were replayed through the production sampler trajctlhook.RunTurnEnd with the real git resolver GitEvidenceResolver(root) (which shells git cat-file -e <sha>^{commit} to confirm each SHA resolves before crediting a phase). At turn k the evidence window carried exactly the phase→commit bindings for phases 1..k — the honest historical replay of the epic as its steps landed. Every row below was produced by the shipped fold; none is hand-authored.

Why a harness and not fak trajctl: the CLI is parked (var _ = cmdTrajctl, no case "trajctl" in main.go) and the turn-end sampler has no live caller, so the spin had to be driven from a throwaway go run harness against the real Go API. That is itself the headline dogfood finding (gap G1/G2 below).

Real ledger rows (fak-trajctl/1)

The declared objective row:

{"schema":"fak-trajctl/1","kind":"objective","objective":{"id":"trajctl-epic-2533","statement":"trajectory control: score anything, steer by curve (epic #2533)","plan":[{"id":"phase-1","title":"objective score ledger spine (#2534)"}, … ,{"id":"phase-7","title":"re-anchor nudge — regime-gated steering (#2540)"}],"budget":{"turns":8},"status":"active"}}

The W3 witnessed-commit-progress score rows (first, mid, and the plateau — trimmed to show the shape; every row is real output):

{"schema":"fak-trajctl/1","kind":"score","score":{"objective_id":"trajctl-epic-2533","value":0.14285714285714285,"method":"witnessed-commit-progress","version":"1","witness":"W3","evidence":[{"kind":"commit","ref":"41e62c0b97dc","detail":"phase-1"}],"unix_millis":1751850000000,"session_id":"trajctl-dogfood-2026-07-07","run_id":"dogfood-1"}}
{"schema":"fak-trajctl/1","kind":"score","score":{"objective_id":"trajctl-epic-2533","value":0.8571428571428571,"method":"witnessed-commit-progress","version":"1","witness":"W3","evidence":[{"kind":"commit","ref":"41e62c0b97dc","detail":"phase-1"}, … ,{"kind":"commit","ref":"f649954e4d84","detail":"phase-6"}],"unix_millis":1751868000000,"session_id":"trajctl-dogfood-2026-07-07","run_id":"dogfood-1"}}
{"schema":"fak-trajctl/1","kind":"score","score":{"objective_id":"trajctl-epic-2533","value":0.8571428571428571,"method":"witnessed-commit-progress","version":"1","witness":"W3","evidence":[ … 6 commit refs … ],"unix_millis":1751871600000,"session_id":"trajctl-dogfood-2026-07-07","run_id":"dogfood-1"}}

The witnessed-progress curve, turn by turn:

turn 1: value=0.1429  evidence=1
turn 2: value=0.2857  evidence=2
turn 3: value=0.4286  evidence=3
turn 4: value=0.5714  evidence=4
turn 5: value=0.7143  evidence=5
turn 6: value=0.8571  evidence=6
turn 7: value=0.8571  evidence=6   ← phase-7 (nudge #2540) has no commit; curve holds at 6/7

The curve fold + the regime-gate decision

== curve fold (schema: fak-trajctl-curve/1) ==
objective=trajctl-epic-2533 status=active signal=HEALTHY latest=0.8571 delta=+0.0000
detail: progress 0.86 (delta +0.00)
method=witnessed-commit-progress version=1 points=7
  curve: 0.143 0.286 0.429 0.571 0.714 0.857 0.857

== regime gate decision ==
decision=ABSTAIN reason=healthy-curve (signal=HEALTHY, no divergence)

The regime-gate decision was a verified abstention. The witnessed curve rose monotonically 0.14 → 0.86 and then held steady with no competing stall/drift/overrun evidence, so classify() returns HEALTHY. The nudge doctrine (#2540, #2533) is that a HEALTHY curve is never nudged — mid-trajectory intervention degrades a high-success session (arXiv:2602.03338) — so the gate withholds. This is the “at least one regime-gate decision (nudge or verified abstention)” the issue requires, and it is honest: the abstention is forced by the evidence, not chosen for convenience.

What mis-scored / the blind spots the dogfood surfaced

• A phase is credited on SHA-existence, not semantic match. The W3 commit scorer credits a phase the moment any bound SHA resolves to a real commit object; it never checks that the commit actually implements that phase. A resolvable-but-wrong SHA (or a trailer pointing at an unrelated commit) would witness the phase just the same. The witness rung is “the commit exists,” not “the commit does the work.” Adjacent to the calibration (#2566) and anti-gaming (#2568) children.
• The STALL (W2) path was never exercised. The activity-divergence scorer only fires when session-audit rows are fed into the window; the dogfood fed none, so STALL was structurally unreachable and the HEALTHY verdict never had to out-rank it. The dogfood proves HEALTHY→abstain but does not prove DRIFT/STALL→nudge (the nudge actuator does not exist yet — #2540).
• The curve plateau reads as HEALTHY even though the objective is unfinished. At 6/7 the spine is genuinely incomplete (the nudge is unlanded), yet a flat top curve with no divergence signal classifies as HEALTHY. “Steady near the top” and “steady but stuck below done” are the same signal without a completion/expectation reference. Not wrong per the pinned rules — but a real interpretive blind spot for a consumer.

The honest miss list (what could NOT be exercised end-to-end)

1. No live producer. trajctlhook.RunTurnEnd/RunCompaction are called only by tests; nothing in a running session samples the curve at turn ends. → filed as #3129.
2. No live source of phase→commit bindings. Even a wired sampler would score W3=0 forever, because nothing assembles WindowInput.PhaseCommits from commit trailers / a verify pass; the dogfood hand-built them from git log. → filed as #3129.
3. The CLI is parked. fak trajctl is unreachable from main.go dispatch. → already tracked by #2765 (unpark + wire the regime-gated nudge).
4. The nudge actuator is unshipped. Only abstention is reachable; a real nudge on a degrading curve cannot be witnessed yet. → already tracked by #2540 / #2765.

Filed gap list (the witness links)

• #3129 — the live turn-end producer: wire RunTurnEnd into a session + assemble PhaseCommits from commit trailers (new, filed by this dogfood).
• #2540 — the re-anchor nudge (regime-gated steering) — the actuator this dogfood could only abstain in front of.
• #2765 — unpark the fak trajctl CLI + wire the nudge onto the control bus.
• #2566 — scorer calibration meter (score-vs-witnessed-outcome) — the frame for the SHA-existence blind spot.
• #2568 — the anti-gaming fence (intent-literal rows, steering allowlist).

Reproduce

Declare trajctl-epic-2533 with the seven phases above, bind each phase to its listed SHA, and replay turn ends through trajctlhook.RunTurnEnd with trajctlhook.GitEvidenceResolver("."); fold with State.CurveReportFor. The SHAs are real trunk commits — git cat-file -e <sha>^{commit} confirms each. The package folds are covered by go test ./internal/trajctl/... ./internal/trajctlhook/....