feat(gardener): strategy-tick subcommand for ADR-0081 (#NEW6)

[gHashTag]

Summary

Implements the orchestrator side of ADR-0081 — Unified Experiment Loop.

Adds tri-gardener strategy-tick subcommand. Pure decision logic over a (experiments, workers, bpb_samples) snapshot.

tri-gardener strategy-tick           # dry-run, print decisions as JSON
tri-gardener strategy-tick --apply   # placeholder, live writes in PR-2

Three decision steps

Independent and idempotent — each tick can run as many times as it wants.

#	Decision	Trigger
1	ResetStaleClaim	Worker last_heartbeat < now - stale_claim_threshold (default 2 min) → orphaned claimed/running rows return to pending
2	SpawnMirror	done champion with final_bpb < gate2_target → enqueue sibling rows at seed, seed+1, seed+2 for 3-seed quorum
3	PriorityBoost	running row whose own best BPB beats the next-best other row by ≥ 0.05 → priority=0

If none fire → emit Noop.

Decision/DDL contract

Decision::action_label() returns one of noop / reset_stale_claim / priority_boost / spawn_mirror / enqueue — exactly the values in the gardener_decisions.action CHECK constraint added in #NEW4. Drift between code and DDL trips decision_action_labels_match_ddl_check.

Tests

cargo test --workspace -- --test-threads=1 → 173/173 GREEN (was 163, +10).

Test	Guards
empty_snapshot_yields_noop	base case
stale_claim_is_recovered	step 1 fires when heartbeat dies
live_worker_keeps_its_claim	step 1 does not fire when heartbeat is fresh
champion_spawns_three_mirrors	train_v2 BPB=1.8921 < gate2=1.90 → enqueues seeds 43,44 (42 already exists)
champion_above_gate_does_not_spawn_mirrors	hybrid BPB=2.19 above gate → no spawn
priority_boost_when_running_beats_leader_by_threshold	step 3 promotes new leader
priority_boost_does_not_fire_when_already_top	priority=0 short-circuits
strip_suffix_handles_rng_and_seed_and_neither	parent-canon recovery
decision_action_labels_match_ddl_check	code/DDL drift guard
decisions_are_serde_round_trippable	JSON serde stable

Honest admission (R5)

• PR-1 ships pure logic + dry-run JSON only. The --apply flag is parsed and explicitly warns that live Neon writes (UPDATE experiment_queue / INSERT gardener_decisions in one transaction) land in a follow-up. The operator can pipe the dry-run JSON into psql manually as an audit-friendly intermediate.
• No bayes-opt yet. Strategy is pure heuristic. Bayesian optimization over (model, h, lr, scheduler) lives in PR feat(gardener): Seed Hunter — power-law fit + leader-relative classifier #67 (feat/seed-hunter) and will be merged into the strategy module after the queue receives 6-12 hours of real samples — exactly per your earlier choice ("оба параллельно").
• Suggesting new experiments (Decision::Enqueue) is not yet wired. The variant exists and round-trips through serde, but the heuristic in decisions_for_snapshot does not yet emit it. Bootstrap of the 41-canon grid is currently a manual INSERT from the operator. Wiring lands together with the live --apply writeback in the next PR.

Cooperative hold

Land after PR #NEW5 (feat/seed-agent) — they share the workspace deps and #NEW5 introduces tokio-postgres with-uuid-1 which strategy uses indirectly via WorkerRow.

R-rules compliance

• R1 Rust-only.
• R5 --apply is honest: prints a warning instead of silently doing nothing.
• R7 Action labels are CHECK-constrained at the DDL level (#NEW4) so audit-trail integrity is type-safe.
• R9 Destructive ops (would be DELETE FROM experiment_queue) are not part of strategy by design — only INSERT and UPDATE-status. Cleanup of stale data is a separate ledger concern.

Anchor: phi^2 + phi^-2 = 3 · TRINITY · SEED→NEON→GARDENER→LOOP.