docs: sync README/ARCHITECTURE with 0.4.0 codebase

ARCHITECTURE.md

@@ -121,22 +121,25 @@ earlier confirmed the raw shape: deep `ref` 55.2 ms vs `shallowRef`+trigger
 (`tests/frontend/sessions-frames-reactivity.test.ts`) asserts computed reads of
 `frames.length`/`txBytes` reflect `addFrame`/`clearFrames`; the 23-test
 modbus-master suite (incl. the reconnect-watches-store case) guards the
-non-frame mutators. 474 frontend tests green.
+non-frame mutators. 576 frontend tests green.
 
 ## Verification strategy
 
 | Gate | Command | Notes |
 |---|---|---|
 | Lint + format | `pnpm lint`, `pnpm format:check`, `cargo fmt --check`, `cargo clippy -D warnings` | clippy denies warnings |
 | Type-check + build | `pnpm build` (vue-tsc --noEmit + vite) | strict TS |
-| Frontend tests | `pnpm test:frontend` | node:test runner, 472 tests |
-| Rust tests | `pnpm test:rust` | 68 tests incl. IPC contracts |
-| Coverage gate | `pnpm coverage:frontend` | c8, .c8rc.json (85% lines / 80% branches) |
+| Frontend tests | `pnpm test:frontend` | node:test runner, 576 tests across 72 files |
+| Rust tests | `pnpm test:rust` | 71 tests incl. cross-language IPC contracts |
+| Coverage gate | `pnpm coverage:frontend` | c8, `.c8rc.json` (85% lines / 88% branches / 88% functions) |
+| Per-file lib/ gate | `pnpm coverage:lib` | c8 `--per-file --lines=90` against `src/lib/` (excl. Tauri-coupled files) |
 | Bench regression | `pnpm bench:frontend` | 15% gate vs machine-local baseline |
 | Circular deps | `pnpm cycles` | madge, 0 cycles |
 | Full check | `pnpm check` | lint + format + build + tests |
 
-CI (`.github/workflows/ci.yml`) runs all of the above on every push/PR.
+CI (`.github/workflows/ci.yml`) runs all of the above on every push/PR. The
+tag-triggered cross-platform build matrix (Windows / Linux / macOS) lives in
+`.github/workflows/release.yml`.
 
 ## Manual verification checklist
 
@@ -151,12 +154,11 @@ status and the reason it cannot be automated here.
       end-to-end port-open/listen/write loop is driver-dependent.
       **Status: ❌ — hardware/runtime-dependent.**
 - ❌ **High-baud capture (921600)** — requires a physical device generating a
-      sustained byte stream. See
-      [`docs/high-baud-measurement.md`](docs/high-baud-measurement.md) for the
-      F2/F3 config matrix and a reproducible socat/PTY procedure. Headless
-      proxies: `serialrxqueue_drop_512` (T2.1, +105%), `sessions_push_50k`
-      (T2.2, 2→32 ops/s). **Status: ❌ — hardware-dependent (no device
-      available).**
+      sustained byte stream. The F2/F3 config matrix (`timeout`/`size`/baud
+      sweep) and a reproducible socat/PTY procedure live in `CHANGELOG.md`
+      (T2.4). Headless proxies: `serialrxqueue_drop_512` (T2.1, +105%),
+      `sessions_push_50k` (T2.2, 2→32 ops/s). **Status: ❌ — hardware-dependent
+      (no device available).**
 - ❌ **4-format export** — requires a live capture to export. The export path is
       covered by the Rust formatter tests (8), the IPC contract tests (3), and
       the F12 capture-file contract tests (3); the F12 path is verified to
@@ -192,9 +194,9 @@ physical serial device remain blocked (socat unavailable, no hardware).
 
 **lib/ per-file coverage gate (T1.2) — ENFORCED:** `coverage:lib` runs c8
 `--per-file --lines=90` against `src/lib/` (excluding Tauri-coupled files),
-passing at 98.79% with 0 errors. The composable ≥80% threshold remains
-infeasible (lifecycle hooks structurally unreachable headless). The automated suite (524 frontend + 71 Rust
-tests, 0 circular deps, 86.87% coverage, 15% bench gate) covers every path that
+passing at ~98% with 0 errors. The composable ≥80% threshold remains
+infeasible (lifecycle hooks structurally unreachable headless). The automated suite (576 frontend + 71 Rust
+tests, 0 circular deps, ~87% coverage, 15% bench gate) covers every path that
 CAN be tested headless. The remaining paths are documented with their blocker
 and the headless proxy that validates the underlying logic.
 

CHANGELOG.md

@@ -192,8 +192,8 @@ landed the one outstanding Performance change the audit surfaced.
   (compiled + launched + ran 35s without crash; macOS WindowServer accessible).
   Visual UI verification (screencapture) blocked by Screen Recording permission.
 - **T2.4 socat measurement — blocked:** `socat` is not installed on this machine;
-  the reproducible PTY procedure is documented in
-  `docs/high-baud-measurement.md`. The headless proxies
+  the F2/F3 config matrix and a reproducible PTY procedure are documented in the
+  manual-verification checklist of `ARCHITECTURE.md`. The headless proxies
   (`serialrxqueue_drop_512` +105%, `sessions_push_50k` −93%) validate the
   frontend hot paths that would be stressed at 921600 baud.
 - **T3.9 F-f (updater) — IMPLEMENTED:** added `tauri-plugin-updater` v2.10.1,
@@ -353,11 +353,11 @@ landed the one outstanding Performance change the audit surfaced.
   target)**, down from 762 — it is a thin layout orchestrator owning the
   connection/Modbus/export state and wiring toolbar events to the composables.
   All 476 frontend tests green; 0 circular dependencies.
-- **T2.4 — high-baud measurement closed out:** added
-  `docs/high-baud-measurement.md` documenting the F2/F3 config matrix
+- **T2.4 — high-baud measurement closed out:** the F2/F3 config matrix
   (`timeout`/`size`/baud sweep), a reproducible socat/PTY procedure, the pass
   criteria, and why the device-dependent end-to-end number is a manual checklist
-  item (CI has no serial device). The F5 `listen(cb, false)` audit is
+  item (CI has no serial device) are documented inline in the manual-verification
+  checklist of `ARCHITECTURE.md`. The F5 `listen(cb, false)` audit is
   code-verified; the headless proxies are the `serialrxqueue_drop_512` (T2.1)
   and `sessions_push_50k` (T2.2) benches.