diff --git a/README.md b/README.md
index 676706d..ad07813 100644
--- a/README.md
+++ b/README.md
@@ -19,6 +19,7 @@ No pip, no npm, no build. Python 3.8+ and a browser.
 
 1. `DESIGN.md` — the design canon (what the game is)
 2. `PLAN3D.md` — the build canon (stack, contracts, lane model)
+   · `docs/MANUAL.md` — the HOW: tool map, asset/site cookbooks, earned gotchas
 3. `THREADS.md` — the append-only lane log. **The project's memory.** Read the
    last few `[I]` integrator entries to know the current state.
 4. `SPRINT12.md` — the current sprint (highest-numbered SPRINTn.md wins)
diff --git a/docs/MANUAL.md b/docs/MANUAL.md
new file mode 100644
index 0000000..3dfab33
--- /dev/null
+++ b/docs/MANUAL.md
@@ -0,0 +1,150 @@
+# HARD YARDS — dev manual & cookbooks
+
+*Written at Sprint 13 integration, distilled from THREADS. The canon docs
+(DESIGN.md, PLAN3D.md) say what the game IS; this says HOW to work on it.
+When this file and a lane's THREADS entry disagree, THREADS wins — fix this
+file.*
+
+## The tool map — what answers what
+
+| question | tool | notes |
+|---|---|---|
+| does everything still work? | `web/world/selftest.html` | must be N/0/0; the count is the per-lane sum |
+| is this yard rig-able by mouse? | `web/world/dev_rigging.html` | N switches yards; `__laneB.switchSite(i)` |
+| what does the sky/storm look like at t=X? | `web/world/dev_skyfx.html` | scrubbable clock; flies debris; warning: it aims you SOUTH |
+| player anims/movement | `web/world/dev_player.html` | |
+| is this site winnable, at what price? | `tools/site_audit/` (audit.html browser, audit.mjs node) | funnel state printed in the header — check it |
+| does rigging save the garden here? | `tools/garden_bench/` + `tools/site_audit/gardenfly.js` | garden outcomes, separation targets, margin rule |
+| per-anchor wind exposure over a storm | `tools/storm_envelope/` | the storm-side second harness |
+| regenerate all canon GLBs | `blender -b -P tools/blender/build_yard_assets.py` | deterministic; `--only <name>` for one asset |
+| job-sheet/invoice design preview | `tools/jobsheet/index.html?v=...` | E's design-ahead handover pattern |
+| deploy to partly.party/hardyards | `sh tools/deploy_hardyards.sh` | ships `web/` ONLY; self-verifies |
+
+## Adding an asset (a prop, a structure, a thing in the yard)
+
+**House rules first (PLAN3D §0):** the runtime asset is always `.glb`, committed
+in `web/world/models/`. Raw `.blend`/`.fbx` never enter the repo (gitignored).
+Two legitimate routes in:
+
+**Route 1 — the factory (canon assets).** Add a builder to
+`tools/blender/build_yard_assets.py`, re-run, commit the GLBs. Nothing is
+hand-edited after export; the factory must produce byte-identical output across
+runs (this is asserted — three full runs, same hashes). This is Lane E's lane.
+
+**Route 2 — vendored external (John's hand-built / MeshGod / library assets).**
+Precedent exists: `models/debris/*` is copied verbatim from the 3D-STORE
+library. Drop the GLB in `web/world/models/`, then wire it (below). If it's a
+one-off, say in THREADS where it came from; if it'll be regenerated, it belongs
+in the factory instead.
+
+**Wiring a prop — the gnome pattern (~5 lines + data, one commit).** There is
+deliberately no generic `props: []` — every prop is a named top-level key in
+the site JSON, wired explicitly, so a site says what it contains. The bike
+(Sprint 12) is the worked example — copy its diff:
+1. Put `bike_kid_01_v1.glb` in `web/world/models/`.
+2. Site JSON gains a key: `"bike": { "x": -7, "z": 9.69, "rotYDeg": 180 }`.
+3. `world.js` loader: one entry in the `Promise.all` + a placement block
+   (position via `heightAt(x,z)`, rotation from the JSON).
+4. Optional collateral: bake `collateral` / `collateral_value` extras in the
+   GLB as the *proposal*; the site JSON's number is *canonical* (the carport's
+   `collateralValue: 180` pattern). Nothing is billed until `collateralFor()`
+   can resolve it — an unpriced collateral returns null and reads as "not
+   scored", never as free… and never as billed either (the gutter was a free
+   failure for two sprints).
+5. Wrecks: same origin, same footprint, mesh-for-mesh swap on
+   `world.wreckStructure(id)` — build the wreck variant from a shared helper
+   so intact/wrecked can't drift (E's `_house_facade()` pattern).
+6. An assert that can actually fail, negative-control-proven (E rebuilt the
+   bike at lean 0° to watch the suite go red before trusting it).
+
+**Conventions the loader reads (bake as GLB extras / node names):**
+- `rating_hint` (float) on anchor nodes — effective failure = `hw.rating ×
+  ratingHint`. Unset ⇒ 1 at adoption; the sim reads the anchor, live.
+- `collateral` (string) on anchor nodes — what breaks when this anchor's
+  corner blows. `collateral_key` on a structure whose priced thing has a
+  different name than the structure (house → "gutter").
+- Named nodes matter: `pickup_anchor` (where carried items sit), the
+  `fascia_anchor_*` family, etc. Check `world.js` adoptAnchor before renaming
+  anything.
+
+**⚠ THE AXIS TRAP (bitten every sprint since 3):** the exporter maps Blender
+`(x, y, z)` → three.js `(x, z, −y)`. A lean toward Blender +Y arrives leaning
+toward three.js **−Z**. Never write a docstring about orientation without
+measuring the exported GLB in three.js coords — E's bike comment lied about its
+own geometry and only a browser-coords assert caught the flip. Dims/tri-count
+verifies CANNOT catch a wrong shape or a flipped lean; look at the render.
+
+## Authoring a site (a "level")
+
+Sites are data: one JSON in `web/world/data/sites/`. Top-level shape (see
+`site_02_corner_block.json`): `id, name, blurb, yard, sun, gardenBed,
+structures, trees, posts, fence, wind, [props…], [separation], _design/_why`.
+Anchors are not listed directly — they come from structures/trees/posts (and
+GLB nodes), typed against the checked `ANCHOR_TYPE` enum in contracts.js.
+
+The authoring loop that works (Sprint 10–13 distilled):
+1. Copy the nearest existing site JSON; keep `_design`/`_why` notes in-file —
+   they're the level's design doc and the repo expects them.
+2. `validateSite` runs at load and fails loud (types, wind block, enum). Boot
+   with `boot({site:'your_site', bank:N})` to jump straight to it.
+3. Wind personality lives in the SITE (venturi axis/gain/throat, tree
+   shelters), storm personality in the storm def. **The venturi axis is site
+   GEOMETRY (a line, mod π), not a wind heading** — reconciled twice in
+   THREADS before that sentence earned its place.
+4. Every trap must be: visible before commit (the enum gives it its pre-rig
+   read), priced (collateral), and mechanically real (rating_hint) — "the
+   price is the trap" vs "the steel is the trap" is a ruling, not an accident.
+5. Run the gauntlet before calling it a level: `site_audit` (winnable lines at
+   $80, funnel ON — check the header), `gardenfly`/`garden_bench` (does
+   rigging separate held-vs-bare? pin a `separation` block if this site makes
+   that promise), `storm_envelope` (per-anchor exposure), then **a cold
+   playthrough by someone who didn't author it** (Lane D's job) — Sprint 11's
+   soft-lock shipped behind 296 green tests and was found in one minute of
+   play.
+6. The blurb and brief carry real design load (night 3's "plenty to tie off
+   to" is the trap's cover story). Client/brief/pay live in week.js NIGHTS;
+   a site without a night is only half real (`boot({site})` seeks the night).
+
+## The earned gotchas (each cost a real sprint-day; receipts in THREADS)
+
+- **Measure through the real chain.** Any garden/load number that didn't go
+  through `commit → attach → skyfx` is suspect: session-level `commit()`
+  bypasses attach and scores every rig as bare (D, S13); a 12 s pre-settle
+  skews SailRig's internal storm clock (B catching A, S13).
+- **Site wind is built in ONE place**: `windForSite()` in weather.js. Three
+  harnesses independently rebuilt it wrong (funnel off ×2, frozen tree sway).
+  Import the helper; never hand-wire venturi/shelters in a tool again.
+- **Margin rule:** any bench corner within ~15% of its rating breaks in the
+  real game (residual bench-vs-UI under-read, cause unfound). Marginal ≠ PASS.
+- **`world.dress()` needs the importmap** — without it the bare-specifier
+  throw is swallowed by callers' try/catch and every ratingHint silently
+  becomes 1.0 (no carport trap, invincible house).
+- **`setHardware()` refuses over-budget quietly** — check `.ok` or fly with
+  `budget: 9999` in tools (this ate four of C's measurements).
+- Documentation cannot fail: an unchecked enum, a JSDoc contract, a comment
+  about geometry — if nothing asserts it, expect it to be wrong someday.
+- Two harnesses, one number. When they disagree, find the variable before
+  tuning anything; the variable is usually the harness.
+- One clone per lane; ports 8823 (integration) / 8824–8828 (A–E) via
+  untracked `.claude/launch.json` (copy the example).
+- Hidden tabs throttle rAF dead: drive with `SHADES.step(dt)` / `.render()`;
+  hidden-canvas boot needs one resize before projections are sane.
+
+## Design canon quick-reference (ruled, don't relitigate casually)
+
+- Business: HARD YARDS. ABN stays sequential (never make it realistic).
+- Carport $180 / gnome $25 / gutter $90; bike deliberately unpriced until the
+  sim can knock it over (tripwire assert enforces this).
+- The wild night sells no clean line — the pyrrhic win is intended. Night 5's
+  garden is beyond saving BY DESIGN and the brief says so. GARDEN_DRAIN does
+  not move. 18–45 m² is an availability floor, not a ceiling.
+- Hail is why sails matter (steep, blockable); rain walks under cloth (don't
+  reopen the rain-angle argument).
+
+## Where knowledge actually lives
+
+READ ORDER for anyone new: README → DESIGN.md → PLAN3D.md → THREADS from the
+last `[I]` entry backwards until you have context → the current SPRINTn.md.
+THREADS is append-only memory, not a reference — when you learn something from
+it the hard way, promote it into THIS file so the next person reads one page
+instead of 300 entries.