fix(prompt): informational asks are not mutation work (#884)

[justrach]

What changed / Why

A default turn treated “go through the codebase and summarize what it does” as implementation work: todos, fan-out, a full test run, and a separate citation pass. That is a policy bug, not tool latency — mutation completion language in the default prompt outranked “match verification to the ask.”

This gates todo, fan-out, tests, and citation-hunting on summarize/explain/inspect verbs and the absence of a requested change. Mutation tasks still require read-before-edit, root-cause fixes, and in-project verification. The harness records the selected intent on the JSONL trace and, after six model rounds on an informational turn, injects one checkpoint asking whether the model can answer now. Lean -p fake_done does not bounce a completed summary.

Evals that would have caught this:

• Classifier unit tests (informational vs summarize-then-fix)
• Prompt pins for the default and lean branches
• Three scripted harness_behavior cases (policy present, checkpoint fires, mutation skips it)
• readonly-summary graff-eval trap: fail if tests/builds ran or sources were edited

Fixes #884.

Verification

• zig build test -Dtest-filter=884 — pass
• zig build test -Dtest-filter=golden — pass
• zig build test — 2227 pass, 1 skip
• scripts/eval-tier1.sh via pre-push — green