v0.3.0 — trust-layer repositioning, mission prune, engine diet, verifier hardening

Trust-layer repositioning: standalone verification, mechanical gates for the shipped surface, an instruction-budget diet, and a pruned mission catalog.

Added

• fleet-verify CLI + GitHub Action — scripts/fleet_verify.py replays every verification layer (findings re-quote, blind-fix anti-anchoring, archive manifest + sha256, fleet-outcome gates, trace schema) against one .fleet/runs/<run_id>/ directory, from this framework or any agent's artifacts. Truly-absent artifacts SKIP by name, but a missing manifest (or a manifest naming evidence a layer then can't verify) is a hard fail; an empty or manifest-less run dir exits 2 (empty verification is not success). Ships with a composite action.yml for PR-side gating and a mutation guard (fleet-verify-empty-run-not-success).
• reviewer_mode in fleet-outcome — enum (cross-vendor-structural | same-vendor-instructed | single-process-instructed) enforced by the validator: invalid values fail, absence warns. Spec in references/fleet-outcome.md; example fixture records cross-vendor-structural; mutation guard (reviewer-mode-enum-guard-off).
• Shipped-mission evidence gate — validate_mission_promotion.py --require-shipped asserts every mission in skills/ carries the progress + readiness + archive triple, resolving ledger names through the mission registry (adversarial's arch-build-* prefix included); wired into validate-all.sh with zero escapes. A documented --known-gap hatch exists but is unused.
• Trigger-loaded engine references — engine-autonomy.md, engine-workers.md, engine-review.md, engine-recovery.md under autonomous-fleet-core/references/.

Changed

• engine.md 944 → 264 lines — always-read core capped at 300 lines by check_instruction_budget.py (docs/instruction-budget.json ratchet); all doctrine moved verbatim into the four trigger-loaded references; every deep link retargeted.
• README repositioned — leads with the verification substrate, answers the single-agent-first critique, documents fleet-verify, and fixes the dead Orca link (diggerhq/orca → stablyai/orca).
• Architecture/menu counts, plugin description, and guide/docs-site catalog updated to 12 exploratory + 6 archived; test suite now 72 files / 1523 tests.

Removed

• Six exploratory missions parked to docs/exploratory/missions/archive/ — release-document, devex-audit, landing-page-convergence, product-framing, security-cso-audit, legacy-rebuild — for overlap with the shipped surface or missing enforcement machinery. Un-parking requires the evidence triple plus a written differentiation case (see the archive README). gstack-quality campaign is now an archived stub; active gstack community-preflight slugs reduce to browser-qa-fix + incident-investigate.

CI / infrastructure

• Restored main CI to green (red since 2026-06-24): fixed a ShellCheck SC2034, made the skill-creator install restore the pinned skills-lock.json, and made ~20 environment-dependent tests hermetic (fake authenticated gh + git identity shim, sandbox-availability guard, ImageMagick for banner normalization).

---

Verification: validate-all green, 100% line coverage of scripts/*.py, 72/72 mutation guards caught. Full changelog: CHANGELOG.md.