research(session-lineage): persist VM decision — Python memo + B=1500 (#54)

[Skywind5487]

Result

The #54 resolver gate is complete.

Selected production shape: ranked sequential indexed point traversal + query-local Python node -> resolved_root memo/path compression.

Selected global work budget: B = 1500 successful row lookups.

Production implementation is tracked separately in #62. This remains a research/evidence PR; it does not modify production session-search code.

Durable evidence

• research/session-lineage/benchmark/FOCUSED_GATE.md — final decision record;
• research/session-lineage/benchmark/results/focused_vm_gate_20260809.md — exact e2-micro receipt, tests, focused performance CSV, safety/B CSV, package SHA;
• python_memo.py — hardened selected research candidate;
• focused_scenarios.py — shallow normal + depth14/size15 compatibility fixtures;
• focused_vm_gate.py — production-VM focused runner;
• tests/test_focused_gate.py — exactness/reuse/budget contracts.

Existing TEMP/Fixed/broad-gate/profiler artifacts remain historical/reference evidence. The old "Pure TEMP vs Fixed3 only" finalist framing is superseded.

Focused VM result

Accepted run source head: 444c161218b00166bda73f2ec5a21f250e2049bc.

The runner used a clean worktree, disposable synthetic DBs only, a 35-second CPU precondition, and passed 11/11 focused/contract tests.

Median timings:

workload	no memo	Python memo	TEMP ref	Fixed3 ref
normal depth0	0.037460 ms	0.038469 ms	0.294199 ms	0.297606 ms
normal depth1	0.051469 ms	0.053704 ms	0.335803 ms	0.320288 ms
depth14/size15 worst-ranked	0.592984 ms	0.124840 ms	0.564849 ms	0.423345 ms
depth14/size15 full-consume	0.619249 ms	0.114357 ms	0.519640 ms	0.386486 ms

The memo cost on shallow normal paths is ~0.001–0.002 ms while reducing the real compatibility extreme by ~4.75x–5.42x. Full-consume work drops deterministically from 120 to 15 successful row lookups.

B=1500

Previously established legitimate frozen-corpus envelope:

C <= 300   -> work <= 554
C <= 1000  -> work <= 1254
real 15-node worst ordering -> work <= 120

B=1500 leaves ~20% headroom above the hard C<=1000 envelope. The pathological e2-micro curve contains non-monotonic quota stalls, so the budget is selected from the logical work envelope + modest margin, not from a fitted wall-time cliff.

Runtime caveat

The saved focused-run receipt used /usr/bin/python3 linked against SQLite 3.40.1 rather than a Hermes-managed runtime. This does not reopen the memo-vs-no-memo mechanism decision; #62 still requires normal Hermes-runtime tests before production merge.

Production semantic handoff

#62 must not reproduce the current generic-parent mistake. The DB winner path and tool-layer current/title exclusion must share one positive compression-continuation root meaning, keeping branch/delegation/tool parent relations separate and respecting the Aug-7 foreign-marker correction documented in #58/#46.

CI note

Research-diff ruff/ty passed. Repository-wide failures observed during this PR were in unrelated dev files/tests (including existing mem-trim and Windows-footgun/lint findings); this PR intentionally does not modify unrelated production code to make global checks green.

Status

• Session-lineage resolver research lane: validate always-memo + B/cycle receipts #54: completed/closed;
• Production integration (BLOCKED on #54): session-search compression-lineage resolver #62: production integration open;
• [research] Reconstruct ChatGPT→Hermes DB import/merge from archived chats + fields #60: independent ChatGPT import/merge provenance archaeology; may refine historical weighting, not reopen this gate by default;
• do not auto-merge this PR.

Refs #54 #58 #60 #62.