feat(search): use compression-lineage memo resolver (#62)

[Skywind5487]

Summary

Implements #62 using the decision from #54 / PR #55:

• replace generic-parent recursive winner lineage handling with a query-local Python node -> resolved_root memo/path-compression resolver;
• enforce a global successful-row-lookup budget of B = 1500;
• preserve existing candidate generation/ranking routes and defer expensive hydration until winners are known;
• make tool-layer lineage/current/title handling use the same compression-lineage semantics;
• fail closed on missing parents, positive cycles, and budget exhaustion;
• keep branch/delegate/tool parent edges separate while allowing legitimate compression continuations, including foreign inherited markers.

Production lineage predicate

A child continues a compression lineage only when all supported conditions hold:

• child.parent_session_id = parent.id;
• parent.end_reason = 'compression';
• child.source != 'tool';
• _branched_from does not point at this parent;
• _delegate_from does not point at this parent.

Foreign/stale markers pointing elsewhere do not disqualify a legitimate continuation.

Implementation

• hermes_state_lineage.py
  • production resolver with query-local memo/path compression;
  • one logical read snapshot;
  • local cycle detection;
  • B=1500 global successful-row-lookup bound;
  • search_session_winners() override preserving existing unicode61 / CJK LIKE / trigram candidate ranking and source-priority ordering;
  • diagnostics for lineage work and bound hits.
• hermes_state.py
  • wires SessionDB to the new search mixin.
• tools/session_search_tool.py
  • routes _resolve_lineage() through the shared production resolver;
  • keeps generic-parent fallback only for older/mock DB objects;
  • title discovery now obtains a lightweight first-message anchor instead of loading the entire transcript, satisfying the existing bounded-hydration contract.
• tests updated/added for positive compression continuation, branch/delegate/tool separation, foreign markers, malformed config, missing parents, cycles, budget exhaustion, memo reuse, source priority, current/title exclusion parity, and candidate routing.

Validation

VM validation on the implementation worktree:

#62 focused tests:        15 passed in 3.97s
session-search sweep:     15 passed in 3.56s

The focused mechanism benchmark from #54 / PR #55 was run on the same VM hardware and selected Python memo/path compression over no-memo/TEMP/fixed-depth references. The benchmark receipt used /usr/bin/python3 + SQLite 3.40.1; this PR does not reopen that algorithm decision. Production Hermes may use a separately patched owner runtime, while this acceptance run exercises the repository code path on the target VM.

Evidence / references

• Session-lineage resolver research lane: validate always-memo + B/cycle receipts #54 research gate / decision
• PR research(session-lineage): persist VM decision — Python memo + B=1500 (#54) #55 raw focused VM evidence
• [research] Separate ChatGPT-import lineage provenance from Hermes-native compaction lifecycle #58 compression-lineage semantics
• Production integration (BLOCKED on #54): session-search compression-lineage resolver #62 implementation contract

Non-goals

• no TEMP-table resolver;
• no fixed-depth production resolver;
• no persistent lineage-root/materialized schema;
• no [research] Reconstruct ChatGPT→Hermes DB import/merge from archived chats + fields #60 provenance archaeology;
• no candidate-ranking redesign.

Closes #62