Gauntlet automatic: full applicable assurance with experimental selectivity - #54

Scope

Upgrades Infinity Gauntlet / Process Assurance as one stacked commit over Council v3 while restoring its automatic watchdog role.

Production automatic path

• preserve the ten canonical typed operations and legacy single-operation API
• add tools/gauntlet_automatic.py as the production controller
• use AUTOMATIC_FULL by default at release
• evaluate every currently applicable canonical operation
• treat structural budgets as advisory telemetry rather than permission to omit applicable release checks
• continue after the first ISSUE so one release attempt can expose all represented blockers
• scan one integrity-valid task/event/receipt snapshot and emit one aggregate ASSURANCE_ONLY receipt
• verify the internal RuntimeStore receipt→event chain so broken typed-event linkage cannot manufacture a green result
• retain the low-level selective/budgeted planner only for explicitly named experimental modes

Experimental selectivity

Only SELECTIVE_EXPERIMENTAL and FAST_BLOCK_EXPERIMENTAL may reduce coverage or stop after the first issue. Their receipts explicitly record that coverage reduction was experimental. No minimum-cost, global-optimality, or efficacy claim is made.

Authority and boundaries

• Gauntlet may clear only ASSURANCE
• target-domain clearance remains unauthorized
• coverage is scoped to RUNTIME_STORE_REPRESENTED_HAZARDS; the event-chain guard does not prove that every real-world hazard was represented
• cost_units remain UNCALIBRATED_ORDERING_PROXY, not measured tokens, money, latency, CPU, or scientific value
• no foil_*, Math Foundry, or Mastermind control plane is imported
• Soul code and routing are not changed in this PR

Changed paths

• tools/gauntlet_runtime.py
• tools/gauntlet_automatic.py
• tests/test_gauntlet_planner.py
• tests/test_gauntlet_automatic.py
• skills/infinity-gauntlet/SKILL.md
• docs/specs/GAUNTLET_ENGINEERING_SPEC.md
• docs/specs/GAUNTLET_AUTOMATIC_SPEC.md
• validation/validate_vnext_runtime.py
• validation/validate_showcase.py

Validation

Final clean commit: 40120719b7adda63dc9a91934ed3801529aea385

Exact source tree: 679a71b85b253f603ccec1862c5047973febe340

That exact tree was validated through commit c364bd6b08f14c310790e5f621d675bfc77b6f17 before being rewritten into the one-commit stacked form with the same tree hash.

Validation on the exact tree:

• complete repository suite: 531 tests passed
• focused automatic-Gauntlet tests: 6 passed
• legacy planner/hardening regressions: PASS
• Research Orchestrator + Process Assurance public invariants: PASS
• vNext typed-runtime validator: 40/40 passed
• public showcase validator: 33/33 passed
• Ruff lint and Python compilation: PASS
• Research software validation run 33201057897: PASS
• Runtime portability run 33201058158: PASS
• Security gates run 33201058019: PASS
• CodeQL run 33201058018: PASS

Stacked validation rule

Base is Council v3 at 7fcfd259a6b103a464061c1d30f2a48fa1ff7f52. If that base changes, this validation is stale and the complete suite must be rerun.

Evidence boundary

Mechanical validation establishes automatic scheduling, snapshot reuse, internal event-chain checking, authority separation, and compatibility. It does not establish lower real tokens, lower total cost, higher benchmark scores, complete hazard representation, or better downstream outcomes. Those require a prospective matched FULL_AUTOMATIC / SELECTIVE_EXPERIMENTAL / NO_GAUNTLET / ORACLE_ANALYSIS evaluation with held-out hazard families, escape counts, false blocks, and whole fix-to-green cost.