@oomkapwn/enquire-mcp 3.11.0-rc.21 → 3.11.0-rc.22

package/CHANGELOG.md

@@ -2,6 +2,10 @@
 
 All notable changes to this project will be documented here. The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and the project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [3.11.0-rc.22] — 2026-06-26
+
+> **TL;DR:** **The rc.20 timing-RATIO hardening itself flaked — re-fixed durably to an absolute floor on the OLD (slow) time + ships the comparison-ready external-audit brief for the v3.11.0 → `@latest` promotion gate.** **(1) TEST-INFRA [release-blocker, recursion of rc.20].** rc.20 converted the 6 wall-clock timing tests' NEGATIVE controls to `quad / linear > 8` ("ratio, robust to runner speed"). But on a docs-only PR's `coverage` job the wikilink dense-`[[a]]` NEGATIVE flaked: `quad/linear = 7.50 < 8`. Root cause: the **`linear` denominator is a sub-millisecond op whose wall-clock is noise-dominated on a contended CI runner** (read ~43 ms there vs ~2.5 ms locally), compressing the ratio below the threshold — the rc.20 fix moved the flake from the POSITIVE budget to the ratio's noisy denominator. Re-fixed durably: the 3 NEGATIVE controls (`wikilink-scan` ×2 + `write` ×1) now assert an **absolute FLOOR on the OLD catastrophic time** (`expect(quad).toBeGreaterThan(50)`) and drop the `linear` baseline entirely. Calibrated empirically: the old functions are ~325 ms (dense `\n`-scan), ~430 ms (lazy-regex unclosed run), ~830 ms (O(n²) slice+concat) on a dev laptop; **a CI runner is only SLOWER, so a 50 ms floor can only fail if the runner ran the quadratic 6–16× FASTER than a laptop (impossible)** — and load pushes the quad's time UP, never below the floor. The POSITIVE budgets (rc.20's generous `< 2000 ms` ceilings) were already robust and are unchanged. **(2) DOCS — comprehensive comparison-ready external-audit brief** (`docs/audits/AUDIT-REQUEST-FULL-FROM-SCRATCH-v3.11.0-rc.21-2026-06-26.md`, targeting the `v3.11.0-rc.21` tag) for the promotion gate. Redesigned from the rc.17 brief after a 3-report Workflow analysis of what 4 LLM auditors did well/badly: **obligation-GATING** tables (the per-always-on-tool sink-trace + the prior-finding re-probe cap the score at 2/5 if unfilled — so a "ship/clean" verdict is expensive to claim), **named worst-case shapes** (dense-CLOSED `[[a]]` run, not an unclosed one — the wrong probe 3 auditors used to re-bless the wikilink quadratic), a **wrong-probe 3-point complexity-curve self-check**, **evidence-or-downgrade** (no self-reported effort scalars; empirical requires a pasted harness + ≥3-point timing + stderr-clean run), a **severity rubric anchored to the real rc.18→20 outcomes** + reachability-proof, an **anti-anchoring novel-class tier**, and a fix to the rc.17 brief's own `40 src/*.ts` count drift (it's 33 top-level / 40 recursive). **No `src/` behavior change** (only the `VERSION` constant); **1416 tests unchanged** (3 NEGATIVE-control assertions rewritten in place). **Lesson: a RATIO timing test is only robust if BOTH terms are above the measurement-noise floor — dividing by a sub-ms baseline reintroduces the very flake the ratio was meant to kill; the durable form for a "the old code is slow" NEGATIVE control is an absolute floor on the OLD (genuinely-slow) time, which load can only push further from failing.**
+
 ## [3.11.0-rc.21] — 2026-06-26
 
 > **TL;DR:** **Post-rc.20 re-sweep — 3 confirmed SIBLINGS of the rc.18 fix classes (the session's signature "instance fixed, adjacent sibling missed" pattern), each 3/3 adversarial-skeptic-confirmed.** A focused 4-lens behavioral re-sweep of the shipped rc.18→rc.20 commit (`41dc699`) hunting recursions of the four fix classes found them in adjacent code the rc.18 sweep didn't reach. **Fixed:** **(1) MED — `obsidian_frontmatter_search` value predicates uncapped (input-cap class, rc.18 sibling).** rc.13 capped the `key` arg but left `equals`/`contains` as `z.unknown()` — and the handler `JSON.stringify`s the predicate and string-compares it against EVERY note's frontmatter across the whole vault, so a multi-MB value is a bearer-reachable O(notes × valueLen) CPU/event-loop amplifier (measured ~3.9 s for a 4 MB `contains` over a 2k-note vault). This is the exact class rc.18 closed for `obsidian_full_text_search`, in the VALUE dimension the inventory missed. Fixed at two layers: a `.refine()` length bound (`MAX_FRONTMATTER_VALUE_LEN` = 8 KiB) rejects an over-cap predicate at the schema boundary, AND the per-note `JSON.stringify(arg)` is hoisted OUT of the loop (it is loop-invariant). The `parser-input-cap-invariant` inventory + its detector were extended to recognize the `.refine()` cap form (the value is arbitrary JSON, so `.max()` doesn't apply) and now pin both `equals`/`contains`. **(2) LOW ×2 — fold-offset in the read/snippet path (`replaceLineOnce` class, rc.18 sibling).** `semanticSearch` (bearer-reachable via the always-on `obsidian_search` TF-IDF arm) and `searchText` (`--diagnostic-search-tools`-gated) both computed a snippet offset with `text.toLowerCase().indexOf(term)` and then sliced the ORIGINAL `text` at that offset — but `toLowerCase()` is NOT length-preserving (`İ` U+0130 → `i̇`, 1 unit → 2; final-sigma; …), so a length-expanding fold char before the match drifts the offset, mis-centring the snippet window and miscounting the line number. Both now route through a shared `foldWithMap` / `foldedIndexOf` (search.ts) that returns offsets into the ORIGINAL string; a `tests/fold-offset.test.ts` unit + behavioral test + an anti-pattern inventory guard (no `toLowerCase().indexOf` offset feeds `sliceSnippet`) pins both sites. **Re-sweep verdict:** the CRLF-drop class was independently confirmed CLEAN (no uncovered sibling — the rc.17/rc.19 inventory guard is complete); `filter_frontmatter`'s uncapped value was checked and is NOT the same DoS (it runs only on the bounded fused candidate pool, O(limit × len), not a whole-vault scan). **1410 → 1416 tests** (+6 fold-offset; the cap is structurally pinned via the input-cap inventory, the hoist is behavior-verified by the existing `frontmatter-ops` equals/contains tests). **Lesson: the post-merge re-sweep is non-negotiable — rc.18 closed the input-cap class for `full_text_search` and the fold-offset class for `replaceLineOnce`, but each had a live sibling one module over (`frontmatter_search` value predicates; the read-path snippet offsets); the durable close for the fold-offset class is the shared offset-safe helper, and for the input-cap class the inventory invariant generalized to the `.refine()` form.**

package/dist/index.d.ts

@@ -7,7 +7,7 @@
  * + `McpServer({version})`) and `src/tool-registry.ts` (used in the
  * `vault-info` resource payload).
  */
-export declare const VERSION = "3.11.0-rc.21";
+export declare const VERSION = "3.11.0-rc.22";
 export { main } from "./cli.js";
 export { buildEmbedText, buildMcpServer, formatReadyBanner, prepareServerDeps, type ServeOptions, type ServerDeps, startServer } from "./server.js";
 export { parsePositiveInt, parseQuantizationMode } from "./tool-registry.js";

package/dist/index.js

@@ -40,7 +40,7 @@ import { main } from "./cli.js";
  * + `McpServer({version})`) and `src/tool-registry.ts` (used in the
  * `vault-info` resource payload).
  */
-export const VERSION = "3.11.0-rc.21";
+export const VERSION = "3.11.0-rc.22";
 // Re-exports — preserve the v3.5.x public surface so http-transport.ts and
 // tests don't need to know about the new module layout. The set below
 // exactly matches the v3.5.x `export` declarations: `main`,

package/package.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://json.schemastore.org/package.json",
   "name": "@oomkapwn/enquire-mcp",
-  "version": "3.11.0-rc.21",
+  "version": "3.11.0-rc.22",
   "mcpName": "io.github.oomkapwn/enquire-mcp",
   "description": "MCP server giving AI agents (Claude Code, Claude Desktop, Cursor, ChatGPT, Codex, OpenClaw) persistent long-term memory backed by your local Obsidian markdown vault. Hybrid retrieval (BM25 + ML embeddings + BGE reranker, RRF-fused), HNSW + int8 quantization, agentic RAG (HyDE + sub-question decomposition), GraphRAG-light (Louvain), standalone Obsidian Bases, PDFs + Tesseract OCR. Vendor-neutral memory layer for any MCP-compatible agent. 46 tools, 19 MCP prompts, 1416 tests, signed npm build provenance (SLSA L2), semver-bound, MIT, zero cloud calls during serve.",
   "type": "module",