Prove the governed Concept Rescue vertical locally

[infotradescout]

What changed

• Adds a Concept Rescue vertical over SI's existing checkpoint, interrupt, RETRACT, re-gate, and stale-approval authority paths.
• Exercises three distinct synthetic artifact classes: a creator-facing proof of concept, a pitch-deck-style concept, and a repository/README concept.
• Makes an explicit, validated RETRACT replacement product intent supersede the rejected interpretation; conflicting ADD/MODIFY overrides still cannot turn repudiation into a product ask.
• Invalidates a pre-correction pendingPlan atomically on interrupt and requires any separately supplied replacement plan to validate against the corrected active intent.
• Rejects prohibited/non-goal/superseded work regardless of plan-controlled analysis / read-only labels. buildAuthorized=false exempts only positively identified read-only actions, while filesystem application, process execution, and authorization boundaries remain strict.
• Carries the branch onto current main without rewriting its already-pushed history.
• Removes an unrelated pre-existing luxury-showcase/PolicyGuard evidence diff through an additive revert, keeping that work out of this focused PR without a force-push.
• Restores portable release validation for the current Step-1 and Concept Rescue files: the manifest explicitly reviews the runtime/reference/test additions, generated caches are excluded from release inventory, and Windows symlink controls degrade with a truthful recorded skip when the OS cannot create their fixture.

Why

The Concept Rescue branch was pushed but had no review path and did not contain the latest main merge commit. Full controls also exposed release-manifest drift already present on main plus one unguarded Windows symlink adversarial fixture. A later Objector replay found that RETRACT left the old product intent and pending plan in place, so reapproval restored a prohibited sales-page task. This PR now keeps the corrected intent, plan authority, and dispatch boundary aligned.

Evidence and limits

This PR proves a local deterministic SI authority flow only. It does not prove real-model behavior, production deployment, cross-model equivalence, external model/tool/worker interruption, cross-client behavior, or commercial readiness. The eval declarations were validated and dry-run; no model was called.

Validation

• py -3 -m unittest discover -s skills/selective-intelligence/tests -p "test_*.py" — 39 tests passed
• py -3 skills/selective-intelligence/scripts/eval.py doctor --json — 41 valid cases; declarations_only_not_model_run
• py -3 skills/selective-intelligence/scripts/eval.py controls --json — 83 controls passed
• py -3 skills/selective-intelligence/scripts/release.py doctor --json — ready; 71 reviewed files; zero errors; model behavior not ready
• py -3 skills/selective-intelligence/scripts/eval_runner.py run --dry-run — 41 cases enumerated; explicitly no model call

On this Windows host, the two symlink-construction adversarial fixtures report explicit platform skips because symlink privilege is unavailable; the rest of the control suite passes.