[babaliauskas]

What this changes

Capture promotion now refuses a turn whose reply claims a tool call that never happened.

• Detection (captures/claims.py): the check looks for a first-person claim, in English, that a tool action already happened ("I searched", "I've checked", "my notes search didn't"). Promises ("let me check", "I'll search once you confirm") and bare tool names don't trigger it. A turn with any other sign that it acted counts as backed and isn't flagged: a tool call, a tool result, a requested tool call, a retrieval event, or more than one model call.
• capture promote exits 1 on such a turn and tells you to re-run with --allow-unbacked-claims.
• capture sync skips such turns and counts them on their own line. If an earlier sync already promoted the case, the skip line names the case file to delete, unless the case is marked provenance: reviewed. When --repair refuses a case, it suggests --allow-unbacked-claims. A refused turn also drops out of later turns' reconstructed history.
• Overrides: --allow-unbacked-claims / --block-unbacked-claims and captures.allow_unbacked_claims in evalshift.yaml. The flag wins, then the YAML key, and the default is to block. An allowed case is promoted with a warning and never asserts expected_no_tools. init scaffolds the key commented out.
• History warning: promotion warns when a case's recorded history contains an assistant reply making such a claim with nothing in its turn to back it. Claims are judged per turn, and history stripped of tool rows isn't checked.
• Docs: DOCS.md, llms-full.txt, docs/{agents,configuration,conversations}.md, plus a spec and a plan under docs/superpowers/.
• README: links the getting-started video on YouTube at the top.

Why

A turn that was offered tools, called none, and still replied "my notes search didn't find anything" used to be promoted as ground truth. The made-up reply became expected, and the case asserted expected_no_tools: true, so the golden suite rewarded the hallucination.

Test plan

• pytest passes: 3505 passed. The only local failure was test_online_sync_says_it_ignores_limits_it_cannot_read. That's a macOS temp-path wrapping artefact (Rich wraps the long /private/var/folders/... path at 80 columns), not caused by this branch. It passes with a short TMPDIR, and the pre-push make ci passed that way.
• ruff check, ruff format --check clean
• mypy --strict src/evalshift_cli clean
• CHANGELOG.md updated under ## [Unreleased]
• Docs updated

Upgrade note: older CLIs reject the captures.allow_unbacked_claims key, so bump any CI evalshift-version: pin before adding it to a config.

🤖 Generated with Claude Code