Fix usage waste, silent psychology drift, and lost-work checkpoints

[infotradescout]

What failed

Selective Intelligence had three real contract gaps:

1. Token efficiency governed startup size but did not stop repeated searches, duplicate contexts, overlapping agents, or evidence gathering across the rest of a run.
2. Product-design guidance did not preserve the silent color-theory, consumer-behavior, anti-scam, and fraud-resistance layer as mandatory core behavior.
3. The word “checkpoint” meant only an intent/approval gate. It did not require automatic, non-blocking savepoints, commits, task-branch pushes, receipts, or resume state before timeouts and long operations.

What this changes

• Adds a whole-run usage governor with bounded repository batches, a three-batch decision point, a single evidence ledger, and no duplicate worker ownership.
• Separates intent/authority checkpoints from durable progress checkpoints.
• Requires progress preservation after each coherent slice and before long tests, builds, migrations, handoffs, context changes, or runtime boundaries.
• Adds a portable checkpoint runtime that stages only task-owned files, preserves unrelated work, blocks routine protected-branch commits, records recovery state, and supports task-branch push.
• Restores silent human decision integrity across public, sales, pricing, onboarding, marketplace, payment, trust, and conversion surfaces.
• Makes color theory, consumer behavior, truthful persuasion, anti-scam, and anti-fraud checks explicit.
• Applies the same contract to the canonical package, strict text guide, and generated ChatGPT adapter.
• Adds regression tests so these behaviors cannot disappear without failing release quality checks.

Hard limits

• Maximum 12 text files or 250 KB per large-repository inspection batch before consolidation.
• After three search or inspection batches, act, narrow, checkpoint/resume, or stop with the strongest supported result.
• No overlapping workers for the same question.
• No more than one completed slice or five materially changed files without a durable checkpoint.
• Checkpoint pushes stay on an existing task branch and never authorize merge, release, deployment, migration, or protected-branch mutation.

Proof included

• Progress-checkpoint self-test commits selected work while leaving unrelated changes untouched.
• Protected-branch test rejects routine checkpoint commits to main.
• Push-without-commit test fails closed.
• Core prompt-budget tests assert whole-run usage, durable checkpoint, color, consumer behavior, scam, and fraud contracts.
• ChatGPT adapter test verifies the new runtime files and runs the checkpoint self-test.

This PR is intentionally draft until the repository quality checks complete.