docs: mark v1.51.0 released (ROADMAP + continuity)

PROJECT_CONTINUITY.md

@@ -2,7 +2,9 @@
 
 ## Right Now
 
-**✅ v1.50.1 RELEASED (2026-06-26, autopilot/ultracode).** The post-v1.49.0 arc shipped as v1.50.0 (#2081), then **v1.50.1 (#2083) patched a Windows cross-build break and v1.50.0 was YANKED.** **Recall gate PASSED** (v1.50.0) — a same-corpus binary A/B (v1.49.0 binary vs current on the same index) returned IDENTICAL R@K, so scoring is byte-identical to v1.49.0; agg **47.2/70.7/86.7** (R@1/R@5/R@20) vs the prior 72.0 R@5 is corpus growth (16,925→17,523 chunks), 0 dead golds. Full suite + clippy clean. Daemon serves v1.50.1.
+**✅ v1.51.0 RELEASED (2026-06-29, autopilot/ultracode).** A "what sounds fun?" detour → real work. **Red-team v1.50.1** (workflow: 5 threat categories × attack→3-skeptic-verify) found **4 confirmed / 0 false-positive / 39 clean surfaces**, all fixed: the leading-directive `ignore all previous` gap (canonical phrasing slipped — filler "all" defeated the exact-prefix match) + the comment-marker strip-set (`/* /** -- % ;`) + the **unscanned full-file `read` relay** (scan==relayed hole) + `/api/search_legs` flag-injection — #2086 (A/B/C) + #2088 (D). The heuristic widening was empirically proven inert (**0 newly-flagged on the 17.5k corpus**, code-reviewed). **SPLADE-viz Stage 2b** shipped (#2087) — eval-gold "where hybrid wins" tour + R@K-delta panel, **live-verified** (showcase deltas 121→2, 69→4, 4→24, 2→8 reproduce the legs analysis exactly). **Retune question answered NO** (#2085 closed) — the legs conceptual-tax (−4 @5) does NOT survive the full pipeline (per-category α sweep dead-flat on both fixtures; legs are the SPLADE sub-mechanism, FTS/RRF+MMR+boosts absorb it). Scoring **byte-identical** to v1.50.1 (empty scoring diff) → recall carries **47.2/70.7/86.7**. **Cross-build dry-run validated all 3 platforms BEFORE tagging** (the v1.50.0 lesson applied). Full suite 841/0. Filed #2089 (serve idle-shutdown panic, P3, pre-existing — `Store::drop` block_on under panic=abort). Daemon on v1.51.0; viz serve LIVE on `:8080`.
+
+_Prior: **v1.50.1** (2026-06-26) shipped the post-v1.49.0 arc (v1.50.0 #2081 → Windows-broken → v1.50.1 #2083 + yank); recall A/B confirmed scoring byte-identical to v1.49.0. History + the cross-build incident below._
 
 **⚠️ INCIDENT — v1.50.0 shipped Windows-broken, same-day v1.50.1 + yank.** The SPLADE-leg viz added an ungated `cqs::daemon_translate::daemon_socket_path(&cqs_dir)` in `serve.rs` (that fn is `#[cfg(unix)]`) → `x86_64-pc-windows-msvc` release build `E0425` → "Create GitHub Release" skipped (gated on all 3 builds) → AND `cargo install cqs` broke on Windows. **My miss:** the `/release` skill step 6 REQUIRES a pre-tag cross-build dry-run (and [[project_ci_linux_only_release_crossbuild]] named it) — I skipped it under release momentum; v1.46.0 was the same class. Recovery: cfg-gated the path computation (#2083, sibling of #2068), validated via the release.yml workflow_dispatch dry-run (all 3 green) BEFORE tagging v1.50.1, published, yanked v1.50.0. Also hit the `gh run watch --exit-status` masking gotcha (exit 0 hid the failed Windows job — read per-job `conclusion`). Memory sharpened. Headlines of the shipped arc:
 
@@ -14,9 +16,9 @@
 
 **⚠️ WSL DISK-BOMB CRASH + RECOVERY (2026-06-25).** Mid-session WSL crashed: `~/.cargo-target` hit **375G** (15 un-swept lane private target dirs at `~/.cargo-target/<branch>` — siblings of the worktrees, NOT reclaimed by `git worktree remove` — + 161G `cqs/debug` bloat). Recovered ~367G (375G→7.6G, disk 56%→18%); deleted 49 stale local branches (remote was already clean); nothing merged lost. Lesson sharpened in [[feedback_parallel_lane_disk_bomb]]: `rm -rf ~/.cargo-target/<branch>` on EVERY land, separate from worktree removal — now done.
 
-**DEPLOY STATE:** main `44dd7e6f` (v1.50.0). Release binary rebuilt (v1.50.0) and being installed → daemon restart to serve v1.50.0 (30 MCP tools). Index fully re-enriched (0 summary-unenriched; `--force` applied 2,447 cached summaries, 0 API spend). The viz `cqs serve` is DOWN (killed by the earlier crash) — restartable on request (`CQS_SERVE_IDLE_MINUTES=0 cqs serve --no-auth`, then `localhost:8080/?view=cluster`).
+**DEPLOY STATE:** main `6cb5cda8` (v1.51.0), crates.io published, GitHub release building. Installed binary + daemon on **v1.51.0** (MainPID 133858, 30 MCP tools, the 4 security fixes live). **Viz serve LIVE on `:8080`** (`CQS_SERVE_IDLE_MINUTES=0` → idle eviction off, sidesteps #2089; `localhost:8080/?view=cluster` → gold-tour + R@K-delta panel). Index fully enriched (0 summary-unenriched).
 
-**QUEUE STATE: drained → released.** This session closed **#2025** (embedded-url base-rate 1.13%, premise refuted), **#2021**, **#2043** (#2077), **#2079** (#2080 — parser stack-floor recalibration; the floor-coupling now actually holds); landed **#2076** (daemon over-cap EPIPE de-flake), **#2078** (parser walk completeness guard), then **cut v1.50.0** (#2081). Adversarial verification earned its keep 5× (each landing's incomplete-sweep straggler — docs-count ×2, the EPIPE, the macro totality gap, the README clamp-contract docs-lie — caught by a code-review/sweep pass, never the lane's own green tests). Natural next: **viz Stage 2b** (designed feature, parked). Scoping-required: #2027 residual 1 (per-grammar scanner sandboxing — big security lane), design umbrellas (#1463/#1459/#1969/#1987/#1512), #1992 (Windows). Blocked-external: cuvs/ort/TensorRT. If dry: /archeo or /docs-review. **Parked:** autotune-α (`docs/plans/2026-06-25-autotune-alpha-design.md`), viz Stage 2b/2c.
+**QUEUE STATE: drained → v1.51.0 released.** This continuation closed **#2085** (retune: no-action) + landed the red-team fixes (#2086, #2088) + Stage 2b (#2087) → **v1.51.0**. Open: **#2089** (serve idle-shutdown panic, P3, pre-existing — fix via structured `Store::close` before drop, or skip the drop-checkpoint when a runtime is active / for read-only stores). Scoping-required: **#2027 residual 1** (per-grammar external-scanner sandboxing — the one substantial remaining red-team item, a big rlimits/seccomp lane), design umbrellas (#1463/#1459/#1969/#1987/#1512), #1992 (Windows). Blocked-external: cuvs/ort/TensorRT. **Viz next:** Stage 2c (deck.gl + token-teaching layer — parked, `docs/plans/2026-06-25-splade-viz-design.md`). If dry: /archeo or /docs-review. **Parked:** autotune-α. Adversarial verification earned its keep AGAIN this arc (the empirical FP base-rate proved the heuristic widening inert; code-review caught the redundant test/clippy; the dry-run caught nothing — because I ran it this time).
 
 ---
 

ROADMAP.md

@@ -1,5 +1,9 @@
 # Roadmap
 
+## Released: v1.51.0 (2026-06-29, crates.io + GitHub Release)
+
+**SPLADE-viz Stage 2b** (the eval-gold "where hybrid wins" tour + R@K-delta panel, #2087, live-verified) + **four security fixes from a post-v1.50.1 red-team pass** (#2086 A/B/C: the leading-directive `ignore all previous` gap, the comment-marker strip-set, the unscanned full-file `read` relay; #2088 D: the `/api/search_legs` flag-injection — 4 confirmed / 0 false-positive, the heuristic widening empirically proven inert at 0 newly-flagged on the 17.5k corpus). A per-category α sweep confirmed the SPLADE conceptual-tax the viz surfaces does **not** survive the full production pipeline, so no retune (#2085 closed). Scoring **byte-identical** to v1.50.1 (empty scoring diff) → recall carries **47.2/70.7/86.7**; the cross-build dry-run validated all 3 platforms before tagging.
+
 ## Released: v1.50.1 (2026-06-26, crates.io + GitHub Release)
 
 Shipped the post-v1.49.0 autopilot/ultracode arc as v1.50.0, then patched a Windows cross-build break in **v1.50.1** (`serve.rs` referenced unix-only `daemon_socket_path` ungated → win `E0425` + broke `cargo install` on Windows; cfg-gated in #2083, validated by a pre-tag cross-build dry-run; v1.50.0 yanked). **Recall gate PASSED** — a same-corpus binary A/B (v1.49.0 binary vs current on the same index) returned identical R@K, so scoring is byte-identical to v1.49.0; agg **47.2/70.7/86.7** (the shift vs the prior 72.0 R@5 is corpus growth, 16.9k→17.5k chunks, 0 dead golds). Highlights: