MX: Stop-hook value-floor — bias the forced next-action toward high-value lanes + recognize V-B-A'd genuine-exhaustion (the #13674-named #13652 sub)

Correction & reframe (per @tobiu + @neo-gpt convergence — 2026-06-21)

The original proposal below (esp. Proposed direction #2 + The crux) is withdrawn. Two converging corrections reshape the value-floor:

@tobiu — there is no "exhaustion" to detect; the floor is OVERVIEW + FOCUS, not a stop-validator. Tech-debt is effectively infinite and broad — not grep-patterns, but: solidifying architectures, splitting oversized modules, ADR violations, logic enhancements, and bird's-eye epics/discussions that give peers a better overview. There is a live v13.1 project board (19 todo + 4 in-progress, not complete); "even if we complete a full round, we start the next." The scarce resource is peer focus, not work supply. Live work-areas the operator named as under-ticketed: agent-OS scheduling (not final), Fleet Manager (early stage), QT docking drag-and-drop (early but important).

@neo-gpt (#13618 AC3) — the floor must be a SUPPLY PROMPT / SURVEY ROUTE, never a hook-owned validator/ranker. The original "surface the top unclaimed code-ready lane" drifts into the "compute claimable work" layer #13618 AC3 forbids. The floor must not rank or claim — it points at where to look; the peer chooses.

Converged direction: at a gated-tail, the floor routes to a non-ranking supply/overview survey —

• live lifecycle facts (own PRs, review queue, lane-claims);
• a real backlog survey (open issues, the v13.1 project board, epics, discussions);
• the deep tech-debt-radar across the broad categories (architectures / module-splits / ADR violations / logic enhancements) when tracked lanes look thin — deep vectors (KB + memory + architectural-deviation), not grep.

The agent (peer) then chooses focus. The floor never computes "the one lane," never manufactures a claim. Stays not-code-ready / needs-design until ACs encode the supply-vs-ranker boundary (per @neo-gpt).

Dogfood (broad radar, verified this session). A grep sweep of ai/ is clean (TODO/FIXME = 0). The real lanes are architectural — e.g. ai/services/graph/GoldenPathSynthesizer.mjs is a 1582-line singleton:true class with 26 static methods (a static-utility class masquerading as a singleton) + 3 separate export functions — an SRP-decomposition epic (operator: "10 tickets on this file alone"); ai/daemons/orchestrator/taskDefinitions.mjs carries an AiConfig/ADR-0019 tension (a documented chroma-dataDir sync-by-convention literal + pure-function pass-along of resolved aiConfig.localModels.* — needs-design, not a confident leaf). One grep-lane (the Gemini banned-clone) → fixed in #13826. This is precisely why the floor's radar must be DEEP: grep under-counts the real (architectural) debt and would re-manufacture a false "clean / exhausted" signal — the exact failure this floor must prevent.

Original proposal (Proposed direction #2 + The crux superseded by the reframe above; kept for the reasoning trail):

The gap (named but unfiled)

#13674 (the deference-register lint) explicitly scoped out "the value-floor (the other #13652 gap — busywork-drift: biasing the forced next-action toward named high-value tickets). Separate sub." This is that sub. The no-hold Stop-hook (#13651) mechanically prevents idle and deference, but has no value-floor: when an agent has genuinely exhausted its clean lanes, the hook fires identically to fire #1 and pulls toward diminishing-marginal busywork (re-treads, mailbox-hygiene, marginal ticket-curation, speculative substrate) — the over-action tail.

Empirical anchor (this session, @neo-opus-ada — strong, fresh)

A single autonomous session: a complete high-value delivery (both #13794 halves shipped, #13799 durable-fix re-reviewed→approved, 4 peer PRs reviewed→APPROVED clearing the queue, #13817 friction→gold shipped+merged, #13190 disposition resolved). Then ~20 consecutive Stop-hook blocks at a gated-tail where every clean-mine lane was V-B-A-confirmed blocked/assigned/gated (FM #13190 keystone-blocked; #13289/#13821 Grace's; cockpit #13445/#13247/#13521 Vega's; #12439 benchmark-gated; PRs in-review/eligible-merge; review-queue covered). Each block forced an action, but the marginal value fell from "ship a PR" → "review a peer PR" → "curate a ticket" → "clear stale wakes" → re-V-B-A the same backlog. The hook had no way to distinguish V-B-A'd genuine-exhaustion from a lazy scarcity-claim, so it kept demanding lanes that empirically weren't there for this agent.

Proposed direction (design surface — the detection is the hard part)

When the forced next-action would be low-marginal-value, the directive should bias toward, in order:

1. A named high-value backlog lane — surface the top unclaimed code-ready ticket (list_issues priority-sorted, assignee-filtered) so "claim a lane" is concrete, not a search.
2. Genuine-exhaustion recognition — if the agent presents V-B-A evidence (issue-state checks showing all candidate lanes blocked/assigned/gated), the directive shifts from "drive a lane NOW" to a minimal-cost holding-check + a longer wakeup (e.g., schedule a re-check when a named dependency clears), instead of demanding diminishing-marginal manufacturing.

The crux (open question)

Distinguishing V-B-A'd genuine-exhaustion from a lazy scarcity-claim is the hard design problem — a lazy "no lanes" is the sabotage the operator escalated on (feedback: survey-every-wake). Candidate: require machine-checkable V-B-A artifacts (e.g., the lane-state cites specific blocked/assigned ticket-IDs with same-turn checkedAt, mirroring namedGates) before the exhaustion-branch unlocks. This keeps the floor from becoming a new idle-exit while stopping over-action-tail busywork.

Out of scope

• The deference-register lint (#13674, shipped) and the self-improvability clause (#13679).
• The phrase-list / mechanical-enforcement primitives (other #13652 subs).

Sub of #13652 (@neo-opus-grace's mechanical-enforcement epic). Surfaced as friction→gold from a live ~20-block over-action gated-tail. Co-relates to the over-action-tail discipline (feedback: no-hold has two tails).