docs: record the real-home write audit for bun test The isolation contract named `~/.kabmo` and `~/.omp`, and a sandboxed full run proves no suite writes there. It also found the two suites that wrote elsewhere in `$HOME` — a fixed probe directory that was created and then deleted, and a per-run config dir that outlived a crashed run — plus the reason `HOME` cannot relocate `getConfigRoot()` mid-run.
docs: record the split build, the generated stylesheet asset and the DB write path Build-Test-and-Release: two builds and why one cannot split the client without publishing server chunks under `static/`; the shell is a Bun HTML route in development and `dist/client/index.html` in production; the measured bundle figures. App-Shell-and-Layout: the two generated stylesheets are a linked, content-hashed asset, the layer-order reason the link comes first, and the `link[data-theme-css]` repair guard. Runtime-and-Toolchain: WAL with `synchronous=NORMAL` and `writeMany`'s chunked transactions, with the measured per-write costs.
docs(runtime): record the test-database leak audit A full-run tracer found thirteen suites opening the real `~/.kabmo/db.sqlite` through indirect `getDb()` callers, plus two opens left over from fire-and-forget writes that outlived `releaseDb`. Records the suites, the call-time path snapshot in `getDb()`, and the `[test] preload` fallback that keeps late work in a temp file, so the next suite that reaches a store is audited rather than re-discovered.
docs(harnesses): record the title, resume and mode-availability rules Three behaviours changed, and each carried a measurement a reader would otherwise re-break: - a recorded chat is NAMED (agent title, else first prompt) and why the old `Untitled` read was a regression of the placeholder row it replaced; - the resume path is per TRANSPORT, with the `omp-acp` exclusion (its load forks → a second row), the fork re-point for the harnesses it does serve, and the "empty agent id means the same as ours" rule; - which mode toggles a harness may offer, and why the pick must be read by one resolver instead of re-derived. Runtime-and-Toolchain gains the `bun --hot` rule: module-scope state that owns a subscriber list is emptied by a reload while the one-shot init guard is not, so every signal-based publish stops — measured as a frozen `sidebar:status` sequence while direct publishes kept moving.
docs: split the manual into a wiki