[v0.17 PERF-Q] Replace broad store reads with selective lookups

[TheGreenCedar]

Package PERF-Q in the v0.17.0 consolidated remediation program.

Contract

Deliver W6.2 and W6.3 without changing packet/search semantics:

• add raw selective edge access for route metadata, preserving every edge field,
  uncertainty value, and deterministic row order;
• keep the three production get_edges() consumers that genuinely require the
  complete workspace graph (index_freshness, affected-state loading, and
  report-source export) on the broad accessor;
• replace only the route-handler broad scan used by node-details/affected route
  metadata, including canonical/OpenAPI variants;
• restrict indexed-symbol name lookup to the bounded candidate IDs before the
  existing final stable sort;
• decorate-sort-undecorate retrieval candidates and packet citations so path
  resolution and citation rank are computed once while equal/NaN ordering stays
  unchanged;
• batch child-existence checks after symbol sort/dedupe/truncation.

The existing get_edges_for_node_id[s] accessor is not a valid substitute for
the new raw route accessor: it applies trail policy that can clear uncertain or
common-call resolution, confidence, and certainty.

Dependencies and order

ANNO-A / #1697 precedes this layer. PERF-Q precedes ANNO-B / #1698 so the
storage_impl/mod.rs order remains:

ANNO-A → PERF-Q → ANNO-B → SRC-C → STORE

EV-5 / #1701 must land before the accepted PERF-Q candidate because it owns the
schema-v2 verdict counters and the retrieval/orchestrator hotspots touched here.
EV-5's live calibration is currently pending, so PERF-Q must record an explicit
same-publication schema-v2 before/after report; a normal non-live
packet_search_eval pass alone is not the zero-movement proof.

Program boundaries

• The implementation PR uses a codex/* branch, targets the ANNO-A branch as a
  native upper layer, and carries Closes #1661.
• Keep source, package, protected-host, installed-runtime, and publication proof
  distinct.
• Preserve atomic publication, pinned readers, crash recovery, freshness,
  fail-closed evidence, observational status/doctor, accelerated embedding,
  proof-owned cleanup, and serial Cargo execution.
• Do not touch product version surfaces, schema.rs, packet_search_eval.rs, or
  its checked-in baseline.
• Do not replace a genuinely global graph read with a selective query.

Done when

• Raw selective route results are field-for-field and order-for-order identical
  to filtering the broad edge result, including uncertain resolved calls.
• Query-plan/progress proof shows indexed route access and materially fewer VM
  steps on the representative high-edge fixture.
• Batched child presence matches the old per-node behavior with one statement
  per SQLite bind-limit chunk.
• Candidate/citation sort keys are evaluated once and preserve stable equal-score
  and NaN ordering.
• Node-details and affected route metadata remain exact.
• packet_search_eval.rs and its baseline are byte-identical, and the
  same-publication schema-v2 before/after report records zero verdict/count/cause
  movement.

Refs #1179