docs: cover trace invariants on the summary surfaces; fix advisory-tag scope and hosted-gate wording

[babaliauskas]

Cross-repo documentation audit after #35/#36 (siblings: evalshift-server #30, evalshift-client PR to follow). The reference sections — DOCS.md Trace invariants, docs/configuration.md#evaluatorstrace_invariants, the llms-full.txt evaluator block — were already right and consistent. The surfaces that summarise them were not updated. Prose only, plus one docstring and the re-vendored schema.

Missing → added

• Profile tables (DOCS.md, llms-full.txt): a trace-rule violations ≤ column, 0 in every profile (what init writes).
• Report layout (DOCS.md, llms-full.txt, README.md): the "Trace rules broken" panel; the executive summary and the strip's mean score Δ are drawn from blocking evaluators (advisory ones only on a prompt nothing else scored).
• README agent section, docs/agents.md ("four new evaluators", the top-level placement carve-out, a paragraph in "Reading a tool run's report"), and the "tool evaluators are not written at top level" comments in DOCS.md / llms-full.txt now carry the trace_invariants exception.
• Hosted wording (docs/hosted.md, DOCS.md, llms-full.txt, docs/methodology.md, CHANGELOG): the server accepts and stores max_invariant_violations since 2026-10-08 and never re-evaluates it; a run that carries its own policy is answered with the CLI's decision, only policy-less runs get the six-budget re-check. The CHANGELOG's "this CLI release waits on that server deploy" is gone.
• FAQs: a breached max_invariant_violations is never inconclusive (docs/faq.md, DOCS.md, llms-full.txt); a CLI upgrade that added config fields moves config_hash (resume FAQs in the same three files). CHANGELOG mentions the analyze recommendation for source-shared breaks.

Corrections

• The "Overall, by evaluator" advisory tag (from fix(report): advisory evaluators out of the summary; trace-rule causes counted as the budget does #36) applies to every blocking: false evaluator, not only trace-rule entries — CHANGELOG, DOCS.md, llms-full.txt, docs/configuration.md said the latter.
• call_count has min_calls/max_calls; there is no exact key (CHANGELOG, docs/evaluators.md).
• docs/evaluators.md: "they all run on every pair" qualified for trace_invariants; the budget gates blocking entries (also docs/traces.md, which now also notes the imported-traces evaluate error and applies_to).
• applies_to: DOCS.md said "where supported"; the four other families' rows in docs/configuration.md said only "Glob list". All now say accepted, enforced only by trace_invariants.
• llms-full.txt round numbering uses the replay's 1-based convention everywhere (violation round_index noted as 0-based).
• hosted/push.py docstring: six of ten budgets, other four.

Schema

src/evalshift_cli/hosted/bundle_manifest.schema.json re-vendored from evalshift-server #30 (one description string on BudgetResult.denominator). Merge server #30 first: until then the local-only test_bundle_shape.py::test_the_vendored_schema_matches_the_server_export differs from server main by exactly that string (CI skips that test; it needs the sibling checkout).

Checks

ruff check / ruff format --check / mypy --strict: green. pytest: 2581 passed, 1 failed — the vendored-schema test above, for the stated reason only (verified: 0 differences against the #30 schema). tests/unit/test_docs_currency.py green.

🤖 Generated with Claude Code