chore(release): 4.4.0 'Curve'

Version bump across all surfaces — package.json + lock, server.json (x2),
README banner + install commands + a v4.4.0 'what's new' callout — and
CHANGELOG [Unreleased] -> [4.4.0]. Bundles this session: Token-loop C
(engram measure --session), honest cluster-bootstrap CIs on the recall bench
(ranker lift +5.0pp [+1.7,+8.5], excludes 0), query/gen path guard (#92),
hermetic skills test (#137), README front-door reposition, .gitleaks.toml.

Audits: tsc 0; full suite 1175/1175; build green; gitleaks clean; versions
consistent (package/server/README all 4.4.0).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: NickCirv

CHANGELOG.md

@@ -6,6 +6,13 @@ All notable changes to engram are documented here. Format based on
 
 ## [Unreleased]
 
+## [4.4.0] — 2026-06-14 — "Curve"
+
+**Why:** honest whole-session measurement, statistical error bars, and a front
+door that leads with what's true. Token-loop C lands `engram measure --session`;
+the recall bench now ships confidence intervals; the README leads with the
+caveat-free differentiator instead of a number.
+
 ### Added
 - **`engram measure --session [--replay <log>]`** (Token-loop C, replay mode) — proves the
   *combined* whole-session structural reduction by replaying `.engram/hook-log.jsonl` through
@@ -14,6 +21,27 @@ All notable changes to engram are documented here. Format based on
   $ saving), with the whole-session figure disclosed as a same-epoch ceiling. Hook-log gains
   two optional fields — `sessionId` (honest session boundaries) and `command` (Bash only, so
   the shell-grep arm is measured, not inferred). Old logs still parse.
+- **Cluster-bootstrap confidence intervals on the recall bench** (`bench/stats.ts`) — recall@10
+  (per-commit macro) and the ranker's lift over a random-order baseline now report a 95% CI with
+  an explicit "excludes 0 → real" verdict. On engram's own repo: lift **+5.0pp [+1.7, +8.5]**
+  (excludes 0); recall@10 macro 45.1% [36.4, 54.0]. The reusable stats engine the resolve-rate
+  A/B will consume.
+
+### Changed
+- **README front door** now leads with the caveat-free differentiator (bi-temporal mistakes from
+  git reverts) + capacity, demoting the per-file structural % into its caveat; the `LLM cost $0`
+  badge → `LLM calls 0` (true + not misreadable as a cost claim).
+
+### Fixed
+- **`query` / `gen` now fail loudly on a path with no graph** (#92) — exit 1 with "run engram
+  init" instead of silently printing "no nodes" and exiting 0 on a missing/typo'd `--project`.
+- **Hermetic skills test** (#137) — the `withSkills: true` test resolves its default via
+  `ENGRAM_SKILLS_DIR` at call time instead of mining the developer's live `~/.claude/skills`,
+  removing a CI flake.
+
+### Security
+- Added `.gitleaks.toml` allowlisting the synthetic PII-gate test fixture
+  (`tests/fixtures/pii-zoo.json`) — fake tokens by design, not real secrets.
 
 ## [4.3.2] — 2026-06-06 — "Proof" (patch)
 

README.md

@@ -1,5 +1,5 @@
 <p align="center">
-  <img src="assets/banner-v3.png" alt="engramx — the universal context spine for AI coding tools (v4.3.2 'Proof')" width="100%">
+  <img src="assets/banner-v3.png" alt="engramx — the universal context spine for AI coding tools (v4.4.0 'Curve')" width="100%">
 </p>
 
 <p align="center">
@@ -86,8 +86,8 @@ mkdir -p /tmp/engram-demo && cd /tmp/engram-demo && \
   echo "export const buggy = () => null;" > src.ts && \
   git add -A && git commit -q -m "feat: add buggy helper returning null causing form crashes" && \
   git revert --no-edit HEAD > /dev/null && \
-  npx --yes engramx@4.3.2 init . && \
-  npx --yes engramx@4.3.2 mistakes
+  npx --yes engramx@4.4.0 init . && \
+  npx --yes engramx@4.4.0 mistakes
 ```
 
 You should see, within 30 seconds, the **bi-temporal pre-mortem** engram auto-captured from your revert:
@@ -119,6 +119,8 @@ On engramx's own repo the **per-file structural context reduction is ~89%** (89.
 
 Works in 8 IDEs and counting — Claude Code, Cursor, Cline, Continue.dev, Aider, Windsurf, Zed, OpenAI Codex CLI. One install, one graph, every tool benefits. Apache 2.0. Local SQLite. Nothing leaves your machine.
 
+> **v4.4.0 "Curve" — 2026-06-14.** New: **`engram measure --session`** proves the *combined* whole-session structural reduction by replaying your hook-log as an honest per-session break-even-P curve (Token-loop C). The recall bench now ships **confidence intervals** — on engram's own repo the PageRank ranker's lift over random ordering is **+5.0pp, 95% CI [+1.7, +8.5]** (excludes 0 → real); recall@10 macro 45.1% [36.4, 54.0]. Plus: `query`/`gen` fail loudly on a bad `--project`, a hermetic skills test, and a front door that leads with the differentiator, not the number. Still no cost claims — a structural reduction, not a bill saving.
+>
 > **v4.3 "Proof" shipped 2026-06-05 — engram's saving is now real *and provable*.** Run **`engram measure`** in your own repo to see the honest structural context-token reduction on **your** code — every disclosure computed live: it's a ceiling, here's the recall, here's the intercept rate, "structural tokens, not your bill." Run **`npm run bench:recall`** for the reproducible proof that engram surfaces the files a change actually touches (recall@10 33% on engram's own repo, decomposed honestly: candidate generation reaches 43%, the PageRank ranker adds +3.2pp over random-within-candidate; 10.4% blind chance).
 >
 > New this release: a **never-worse gate** on Grep (engram passes through whenever its packet isn't actually smaller — sized to your grep's exact `cwd`/`path`/`glob` scope), **Bash-grep interception** (the shell-only IDEs — Aider, Codex CLI, Cline — now get the call-site packet too), a **sub-agent context broker** (a tight ~100-token ranked slice into each spawned Claude Code sub-agent — the one regime prompt caching can't help), and a **compaction ledger** (a "previously read" list injected at `/compact` so the agent doesn't re-explore). Every number is a measured fact or a labelled bet — **no cost claims**; a *structural* context reduction, not a bill saving (engram's net over prompt caching ≈ 0). 1149 tests. See [CHANGELOG.md](CHANGELOG.md), `docs/COMPARISON.md` (vs the other local code-graph tools — ranking isn't unique, the *combination* is), and `docs/FRONTIER.md`.

package-lock.json

@@ -1,6 +1,6 @@
 {
   "name": "engramx",
-  "version": "4.3.2",
+  "version": "4.4.0",
   "mcpName": "io.github.NickCirv/engram",
   "description": "The context spine for AI coding agents. 9 built-in providers + mcpConfig plugin contract (wrap any MCP server in 10 lines), generic MCP-client aggregator (stdio), pre-mortem mistake-guard, bi-temporal mistake memory, Anthropic Auto-Memory bridge, SSE streaming context packets, dual-emit AGENTS.md+CLAUDE.md. Up to ~89% per-file structural context reduction (codebase-dependent; reproducible bench included). Local SQLite, zero cloud.",
   "repository": {

package.json

@@ -6,12 +6,12 @@
     "url": "https://github.com/NickCirv/engram",
     "source": "github"
   },
-  "version": "4.3.2",
+  "version": "4.4.0",
   "packages": [
     {
       "registryType": "npm",
       "identifier": "engramx",
-      "version": "4.3.2",
+      "version": "4.4.0",
       "transport": {
         "type": "stdio"
       },