Define recovery for an agent name whose discovery coverage remains unresolved

[pengfei-threemoonslab]

Found while implementing #543; deferred rather than treating every missing name as a human declaration.

What was observed

The bounded #543 classifier can distinguish readable names confined to tests/templates from a selectable product declaration. It cannot make that distinction when the parse is capped, manifest scope is ambiguous, or some Python files were not parsed. cli/discovery/signals.py::_parse_python_facts drops read/decoding/syntax failures; a real fixture containing a readable test Agent plus invalid Python has python_parse_truncated: false but workspace_signals.python_file_count < python_file_total.

Doctor recomputes discovery when agent.name is unresolved, but has no --max-python-files option. A higher-cap detect command followed by unchanged doctor would repeat the same cap. The #543 regressions therefore refuse to classify incomplete discovery as test/template-only and do not offer that unusable retry; the broader unresolved case retains existing placeholder recovery.

This leaves a real recovery question: a generic request to replace CHANGE_ME does not tell the caller whether to repair a source, select a project, provide a supported declaration, or ask a person. Dynamic names, unresolved imports and low-quality literals also lack a positively established cause. Setup permissions remain all false; this is not an observed authorization bypass.

Follow-up

• Settle the recovery vocabulary from parser/discovery facts, without matching rationale or exception prose.
• Make any capped retry consumable by the command that reruns the decision; do not create a detect/doctor loop.
• Preserve exact manifest scope and product-backed coding-agent recovery. Do not make all unresolved names human-owned.
• Specify how a discovery answer handles files changing between its reads before claiming generation-consistent evidence; Preserve why agent.name is unresolved through init and doctor recovery #543's invocation-local facts are not a verifier snapshot.

Acceptance

• Incomplete, unreadable and unsupported name evidence produces a truthful next step naming the missing evidence or decision.
• Any offered automated retry makes bounded progress when executed as emitted.
• Paired product/test, capped, invalid-Python and cross-module cases keep ranked actions and compact control consistent.
• No synthetic product identity, broader setup permissions or hidden coverage claim is introduced.

Related: #543 owns the known test/template-only slice; #328 covers failure attribution. No new public field or command is assumed by this issue.