Make packet evidence selection obligation-aware

[TheGreenCedar]

Objective

Improve packet evidence quality for material flow questions by selecting verified source ranges and typed graph receipts against stable semantic proof obligations, without query-token heuristics, repository-specific steering, higher global caps, or a material latency/reliability regression.

Current evidence

• Monolog retrieves the relevant logging code, but packet compilation can omit the handler iteration/dispatch source even when an exact typed CALL receipt exists.
• AutoMapper can retrieve TypeMap.cs, then lose it during bounded candidate resolution.
• Animate can retrieve its source entrypoint, variables, and animation files, then lose them before final packet selection; CSS import targets and class-to-keyframe linkage also need truthful typed receipts.
• Jekyll is the unaffected regression guard.
• Two selector-shaped revisions and one producer-plus-selector revision were rejected and reverted. The next design must move proof authority out of name-shaped carrier predicates.

Required architecture

• Retrieval query IDs remain provenance only.
• Material obligations use typed proof atoms over exact verified source aspects and typed relation receipts.
• Candidate admission, proof verification, and final selection remain separate.
• The selector reuses existing retrieval, graph, source-read, caps, one-drill, and fail-closed machinery.
• No benchmark vocabulary, prompt-token proof heuristics, quoted-path reconstruction, global top-k changes, ranking-weight tuning, or embedding changes.

Acceptance

• Direct exact-binary packets retain the required Monolog, AutoMapper, and Animate evidence; Jekyll does not regress.
• Two exact-binary Monolog agent repeats plus one AutoMapper, Animate, and Jekyll run are at least as accurate as the unchanged baseline while materially reducing agent work.
• No material packet-latency or reliability regression.
• Focused serial --locked checks pass.

Refs #1179