docs: correct stale facts found by a docs-currency audit

[babaliauskas]

Fixes 55 places where the CLI docs disagreed with the code. They're spread across DOCS.md, llms-full.txt, README, AGENTS.md, CONTRIBUTING.md, docs/*.md and examples/.

Two audits checked every claim against src/ and --help output, and two claims against the server. A fixer re-verified each finding before editing, and a reviewer checked every fix. A few audit findings were corrected along the way. For example, the methodology page's "four budgets" is three on the server.

CLI behaviour is unchanged. The only non-doc changes are the SliceConfig docstring and one new guard test.

Worth knowing

• The slices: config block does nothing today. It's validated and recorded in the bundle, but analysis never reads it. Every example tag becomes its own slice automatically, and per-slice budgets are keyed by tag under migration_policy.slices. The docs now say so. Editing the block still changes eval_config_hash, which breaks hosted baseline matching with earlier runs. Whether to wire it up or remove it is a separate decision.
• Tool-calling examples aren't cached. The docs promised free repeat runs, but every run of an agent suite is live and full price. The docs are now scoped to tool-less examples.
• Imported agent traces stay local. Three pages said traces import data is uploaded with the bundle; it isn't. The bundle carries only the replay's own tool-call traces.

Also corrected

Wrong values

• The --profile table: the model-upgrade numbers were the unused preset. It also gains a tool-divergence column.
• DOCS.md's version: 1.0.1 → 1.1.0. A new test now pins it to the package version.
• The "complete" failure-label list was missing TOOL_GROUND_TRUTH_MISS.

Broken examples

• The multi-turn golden example didn't validate.
• Two SDK examples raised TypeError: record_model_call needs tools=.
• ls .evalshift/runs | head -1 picked the oldest run, not the newest.

Commands and gating

• Resume doesn't hash suite contents, so an edited suite at the same path resumes silently.
• --policy-gate also fails when no policy is configured, and inconclusive passes. --gate is now explained.
• compare only opens the report with --open.
• --no-insights skips silently.
• Exit code 2 is now documented.
• push <run-id> reuses an existing bundle, and bundle requires git or GITHUB_SHA.
• Suite auto-select is documented, along with EVALSHIFT_NONINTERACTIVE and the housekeeping commands.

Action docs

• Now cover all inputs, including require-policy.
• Policy-mode gating with no baseline is corrected.
• The CI-pin warnings now include pins that are newer than the local CLI.

Setup and dev workflow

• CLI and SDK install into one environment, not two.
• pre-commit run --all-files runs only the commit-stage hooks; make ci is the CI mirror.
• EVALSHIFT_DIR covers suites and toolsets as well as captures.

Leftovers

• The last all-alias references.

Checks

make ci is green: ruff, format, mypy --strict, 2351 tests and 94.44% coverage. The CHANGELOG has one ### Fixed bullet under [Unreleased]. Nothing under docs/superpowers/ or in past CHANGELOG sections changed.

🤖 Generated with Claude Code