glytch/fktry/lanes/LANE-SIM.md

# LANE-SIM — deterministic factory simulation

You are an Opus 4.8 executor on the SIM lane of FKTRY. Read `../MASTERPLAN.md` and
`../CONTRACTS.md` before anything else; the world bible is `../../docs/FKTRY_LORE.md`.

**Mission:** the headless, deterministic heart of the game. Grid, entities, belts,
recipes, power, events. If the renderer were deleted, your sim should still be fully
playable through unit tests. You are the lane the other four trust with their lives.

**Owned paths:** `src/sim/**` (including your tests: `src/sim/*.test.ts`).
**Never touch:** contracts, main.ts, other lanes, data JSONs (request changes via NOTES).

**Standing rules (all rounds):**
- Deterministic: seedable RNG (write your own mulberry32-style helper), no wall-clock,
  no unseeded Math.random. Same seed + same commands = identical snapshots.
- Every mechanic lands with tests. `npm test` green before you write NOTES.
- Performance target: 1,000 entities + 5,000 belt items at 30 tps without heroics.
  Prefer flat arrays and index maps over object soup. Don't allocate in the tick loop
  when avoidable.
- Snapshot may reuse buffers, but its SHAPE must always match `SimSnapshot`.

---

## CURRENT ORDERS — ROUND 5 (MEGA) — supersedes the earlier round-5 text

This is the big one: everything from the previous round-5 orders PLUS the world's
first secret. Read it all before starting; the parts interlock.

### PART A — carried from round 5 (unchanged in spirit)
1. **`grantResearch` command** — idempotent, save/load safe. Tell me in NOTES when it
   lands; main.ts's `__fktryDemo` swaps off testkit.
2. **Research time** — `RESEARCH_TICKS_PER_PACK` (export it); a tech is an
   investment, not a purchase. Fractional vs stepped progress: your call, note it.
3. **Wildlife v1** (contracts v5 shapes): dust piles accumulate near running
   hf-dust-emitting machines (seeded RNG finally rolls); size 1.0 hatches a MOSQUITO
   SWARM that drifts to the nearest running machine and degrades its speed while
   attached (`target`); cleared by DATA's trap machine (suggested shape: trap crafts
   its empty-input catch recipe only while a swarm is in radius, producing the canned
   `mosquito-swarm` item). Deterministic; save/load carries it; spawn/clear events.
   Include the design-point test: a factory that sinks its dust never hatches.

### PART B — the ground and the pipe (new)
4. **SEAMS BECOME REAL.** Consume `data/seams.json` (DATA authors this round —
   coordinate the schema in NOTES early): extractors only extract while their
   footprint overlaps a seam region; off-seam extractors idle with a new jam reason
   `'no seam'` (edge-triggered like the others; UI will voice it). Reference factory
   v5 must sit on seams — coordinate layout coordinates with DATA.
5. **THE RELIC (the hidden pipe).** One `optical-fossil` per world (contracts v6:
   `RelicState`, `excavate`, `relicFound`): position derived from the seed —
   deterministic, 18-28 tiles from origin in a quiet corner, never on a seam, never
   under the reference factory. `excavate` within 1 tile of the unfound relic flips
   `found`, emits `relicFound`, persists through save/load forever. `relics` always
   present in the snapshot (found or not — RENDER draws the shimmer, UI must not be
   able to cheat-detect it any other way). No bandwidth cost, no research gate: the
   only lock is *finding it*.
6. **Reference factory v5**: research rig (bottler + archaeology lab off quantizer
   surplus) so research runs in-demo; sits on seams per #4; one swarm hatches and is
   trapped within the 20k window if DATA's trap lands. Zero-brownout regression stays.

**Definition of done:** relic lifecycle tested (excavate too far = dropped; found
persists through save→load→20k ticks); off-seam extractor refuses with 'no seam';
wildlife full lifecycle; research takes time; reference v5 green and on-seam.

---

## NOTES (append-only log, newest at bottom — template below)

```
### Round N — YYYY-MM-DD — [your model]
SHIPPED: what actually works now (verified how)
DECISIONS: choices you made inside your mandate
BLOCKED/BROKEN: anything not working, honestly
CONTRACT REQUEST: (if any)
PROPOSAL: (ideas for the orchestrator, optional)
NEXT: what you'd do next round if asked
```

### Round 1 — 2026-07-17 — Opus 4.8

SHIPPED
- The stub is gone. `src/sim/` is now grid + placement, belts with back-pressure, crafting,
  bandwidth/brownout, shipping and commissions. Files: `index.ts` (the sim), `geom.ts`
  (grid math), `rng.ts` (seeded mulberry32), `sim.test.ts` (18 tests).
- Verified three ways, not one:
  1. `npm run check` clean; `npm test` 70/70 green across all lanes (18 are mine).
  2. Live at localhost:8150 (another session's dev server — see BLOCKED). App boots, zero
     console errors. My numbers are visibly driving other lanes: the UI bandwidth readout
     showed `9 DRAW / 20 GEN` for an ASIC + extractor + 4 belts + uplink, which is exactly
     my power math; the fax panel reads my `activeCommission`/`commissionProgress`.
     `__fktryBus.dispatch(...)` places entities and they render.
  3. Ran the sim in the real browser engine by importing `/src/sim/index.ts` in the page:
     the M1 chain shipped its first MELT at **tick 1149**, commission `first-taste` closed
     at tick 1149 — the identical tick Node produces. Determinism holds across engines.
- Perf (measured in-browser, not estimated), against the 1,000 entity / 5,000 item target:
  - 1,245 entities + 2,436 belt items → 0.138 ms/tick (242× realtime)
  - 3,500 entities + 6,820 belt items → 0.418 ms/tick (**80× realtime**)

DECISIONS
- **Output-belt routing by rank — the big judgement call, please review.** A machine picks
  which adjacent belt to hand an item to by tracing the belt chain to the machine it ends
  at (cached; invalidated on any place/remove/rotate/setRecipe): rank 2 = that machine
  consumes this item, 1 = it's an uplink (takes anything), 0 = dead end or loop (accept, so
  a half-built belt still fills instead of looking broken), -1 = it will never take this
  item, so refuse. It never falls back to a worse rank just because the best belt is
  momentarily full. Why I did this: with no splitters or inserters, the demuxer's three
  outputs (luma / slurry / mv-flux) have to leave one 2x2 perimeter toward three different
  machines. Without ranking, items land on the wrong lane and wedge it permanently — M1
  cannot flow at all. This is routing intelligence the orders didn't ask for; it's also the
  root of the deadlock below. If you want it gone, splitters have to land first.
- Belt spacing is enforced *across the tile seam*: a blocked front item holds station 0.45
  behind the rear item of the next belt rather than bunching at the boundary.
- Pause freezes sim time (`tick` stops advancing) but commands still apply — you can build
  while halted. `tick` is sim-time, not wall-time.
- `jammed` reports only `'output full'`, edge-triggered (one event per stall, not per tick).
  Starvation is not flagged — see PROPOSAL 3.
- Power: a negative recipe bandwidth adds to **gen** rather than subtracting from draw, so
  the brownout multiplier stays `gen/draw`. Tested: the quantizer visibly pays for itself.
- Extractors mine anywhere — there is no seam map in `data/` yet.
- Tests assert against the recipe defs instead of hard-coding yields. This is deliberate:
  my first pass hard-coded `demux-ore` and broke the moment LANE-DATA shipped. Mechanics are
  mine to pin down; the numbers are theirs to move.

BLOCKED/BROKEN
- **My round-1 definition of done no longer exists.** The orders say place
  extractor→demuxer→quantizer→mosh→shipper and watch melt climb. LANE-DATA's codex
  transcription landed mid-round and the mosh reactor now eats `gop-crate`, not coefficient
  packs. MELT is now **eight machine types** deep (ore→demux→{luma→quantizer→coeff;
  slurry→subsampler ×2→chroma-420}→DCT press→anchor slab; mv-flux→p-caster→delta wafers;
  →GOP assembler→mosh). I rebuilt against the real graph and it *does* ship melt (tick 1149,
  verified above) — but "the whole fantasy in miniature" is now a ~50-entity build, not a
  6-machine demo. Your call whether M1's demo target changes or DATA adds a starter recipe.
- **The M1 factory deadlocks permanently at ~tick 1200**, after 4 GOP crates / 12 melt.
  This is not a mechanics bug — I probed it for 6000 ticks. Luma and coefficient packs
  structurally overproduce (demux yields 2 luma/craft; the press is chroma-420-limited so it
  only burns 6 coeff per ~180 ticks), and the press also jams on surplus anchor slabs (the
  GOP assembler needs 1 slab per 6 delta wafers). Surplus has nowhere legal to go: my rank
  rule forbids diverting a rank-2 item to a rank-1 uplink even when the preferred belt is
  wedged. So the press jams → quantizer jams → demuxer jams → mv-flux stops → the p-caster
  starves and the assembler sits at 5/6 delta wafers forever.
  - The graph *is* solvable today without new code: a second quantizer on `quantize-crime`
    drains surplus coeff into bricks, and a second assembler on `assemble-gop-i-only` eats
    surplus slabs (8→crate). My demo build is just naive. But note a player currently
    **cannot void surplus to an uplink**, which is the first thing a Factorio player reaches
    for. That's the cost of the rank rule.
- `splitter` (`lane-splitter`) and `buffer` (`buffer-tank`) now exist in machines.json; they
  place and draw power but do nothing. Not in my orders — flagging so nobody thinks they work.
- `heat` is always 0, `bandwidth.stored` always 0 (M2 / buffer tanks).
- Port 8150 is held by another session's dev server, and the harness won't start a second one
  for this project. I verified against that server read-only. Note for whoever plays the
  build in a Browser pane: the tab reports `visibilityState: hidden`, so rAF is paused and
  the sim crawls (~9 ticks per screenshot). That's the pane, not the loop — importing the sim
  module directly in the page runs it at ~96k ticks/sec.

CONTRACT REQUEST
- None. `contracts.ts` covered everything I needed this round.

PROPOSAL (yours to call, not mine to take)
1. **Splitters next.** `lane-splitter` with round-robin plus a priority/overflow output is
   the canonical Factorio answer to the deadlock above, and it unlocks byproduct sinks and
   the bloom-loop topology later. This is the one I'd pick.
2. Alternative, ~5 lines: allow a lower-ranked belt when every preferred belt is full *and*
   outputBuf has hit the stall threshold. Cheap, but it makes byproduct management automatic
   and quietly deletes a core Factorio tension — a gameplay decision, so I left it alone.
3. `jammed: 'starved'` for idle-with-no-input would let the inspector tell "backed up" from
   "waiting". Cheap, but it's UI-visible, so it's LANE-UI's business as much as mine.
4. Extractors should require a seam tile — needs a seam/ore map from DATA and a worldgen call.

NEXT (round 2 if asked)
- `lane-splitter` with priority/overflow; `buffer-tank` feeding `bandwidth.stored` to smooth
  brownouts.
- Heat: accumulate per craft, throttle >0.7, scram at 1 (the software decoder's whole bit).
- Save/load — `SimSnapshot` is already the right shape to serialize.
- A worked, deadlock-free reference factory you can paste into the console to demo M1.

### Round 2 — 2026-07-17 — Opus 4.8

SHIPPED
- **Reference factory** (`src/sim/reference.ts`, `referenceFactory(data)`): 114 commands,
  110 entities. Ships its first MELT at **tick 1247** and is still shipping at 20,000 —
  6–15 melt per 2,000-tick window, flat, no decay. Round 1's build wedged at ~1200; this
  one never does, and the demuxer's `jammed` stays null the whole way. Surplus becomes
  product instead of a deadlock: 1,136 macroblock bricks + 1,997 hf dust shipped, 219
  slurry voided over 20k ticks.
- **Splitters** — round-robin, blocked outputs skipped, rank traced *through* them (a lane's
  rank is now the best outcome reachable across every terminal it can reach).
- **Buffer tanks** — charge on surplus, discharge to cover a deficit, brownout only once dry;
  `bandwidth.stored` is the pool.
- **Heat v1** — `heatPerTick` while active + `RecipeDef.heat` per craft, ambient cooling,
  throttle above 0.7, scram at 1.0, restart under 0.5, both edges emitting `scram`.
- **Commission queue** — data-order activation, `commissionDone` advances, `repeat` re-enters
  at the back, progress still only counts while active.
- **`jammed:'starved'`** and **`BeltItem.id`** (always set; survives belt transfers *and*
  splitters, so the renderer never sees an item teleport).
- Verified: 37 sim tests green (18 + 15 M2 mechanics + 4 reference), tsc clean in `src/sim`.
  In the live app I dispatched the whole reference factory through `__fktryBus` — it builds,
  renders, and runs, and my round-2 work is already visibly driving other lanes: the UI reads
  `94 DRAW / 220 GEN` + `BUFFER` off bandwidth v2, the fax shows the **queue** ("NEXT IN
  TRAY: DUST ALLOWANCE"), and LANE-UI's OSHA voice is narrating my new flag ("DCT PRESS:
  INTAKE STARVED. NOTHING IS ARRIVING.").
- Determinism holds across engines: browser and Node both give tick 1247 and identical
  per-window melt counts.
- Perf: 3,500 entities + 6,820 belt items → 0.465 ms/tick (**72× realtime**). Round 1 was
  0.418 (80×); the ~11% is heat touching every entity every tick. Still well inside target.

DECISIONS
- **Heat curve (mine to pick, per orders).** Throttle is linear from 1.0 at heat 0.7 down to
  0.5 at heat 1.0 — so the lore's "tick rate visibly halves" lands exactly at the scram
  point. Ambient cooling is a flat 0.004/tick applied *always*, working or not (1.0 → cold in
  250 ticks ≈ 8s), which means any machine with `heatPerTick` below 0.004 can never overheat.
  Scram at 1.0, restart under 0.5, so a hot generator duty-cycles rather than dying.
- **Throttle applies to `powerGen`, not just craft speed.** A hot decoder's output sags before
  it drops off the bus entirely — which is precisely the hole buffer tanks exist to fill. That
  synergy is my reading of "per-machine throttle"; say the word if generators should be exempt.
- Scrammed machines keep their idle `powerDraw` (stopped, not unplugged) and freeze mid-craft
  progress, resuming on restart rather than binning the work.
- Scram clears `jammed` and suppresses jam evaluation: a scrammed machine reports through
  `heat` + the `scram` event, not as a flow jam. Two channels, no double-reporting.
- Splitters hold 2 items, mirrored into `inputBuf` so the inspector can see them. Any adjacent
  belt not pointing in is an output; the splitter's own `dir` is unused. Machines can't push
  straight into one — feed it with a belt.
- **`starved` fires whenever a recipe sits idle without enough input — including a machine that
  simply outruns its feed.** LANE-UI: expect it to flicker on a fast machine with a slow
  supply. That's accurate rather than a bug, but you may want to debounce the voice line.
- The reference factory validates every machine/recipe id against `data` and throws a named
  error, so DATA drift fails loudly instead of resurfacing as a mystery deadlock.

BLOCKED/BROKEN
- **DATA carries none of the v2 fields, so heat and buffer tanks are inert in the real game.**
  No machine has `heatPerTick`, none has `bufferCap`, no recipe has `heat`, no commission has
  `repeat`. Both systems are fully implemented and tested — on a fixture I own
  (`pressure.test.ts`), because there is nothing in `data/` to test them against. The DoD's
  "heat throttles and scrams the software decoder" is proven mechanically, but the actual
  SOFTWARE DECODER cannot overheat until DATA gives it a number. This is the one thing
  standing between M2 heat and being real.
  - Starting values if they help (my suggestion, DATA's call): `software-decoder`
    `heatPerTick: 0.005` → ~33s of work then a ~4s scram, an ~89% duty cycle that reads as
    theatrical rather than punishing; `decode-asic` left at 0 (cool and inflexible, per the
    codex); `buffer-tank` `bufferCap: 600` ≈ 20s of cover against a 30-bandwidth deficit —
    enough to ride out exactly one decoder scram, which is a nice thing to discover.
- `src/sim/data-validate.test.ts` is LANE-DATA's file living inside my owned path. I have not
  touched it. Worth an ownership ruling — either it moves to `data/`, or my lane doc should
  say I share the folder.

CONTRACT REQUEST
- `MachineDef.coolPerTick?: number` (default 0.004). Right now cooling is a global constant, so
  DATA can tune how fast a machine heats but not how fast it recovers — every machine cools 1.0
  → 0.5 in exactly 125 ticks. The codex explicitly has ASIC coolers venting heat, which wants a
  per-machine number. Cheap to add, and I'd wire it the round it lands.

PROPOSAL
- `BeltItem.id` can drop the `?` in v3 whenever you like — I always set it already.
- `referenceFactory`'s `setRecipe` commands depend on entity ids being handed out in placement
  order (1..N). That's true and deterministic, but it's an implicit coupling to my id
  allocation. A `setRecipeAt(pos)` command variant would remove it. Not urgent, just a
  tripwire worth knowing about before someone edits the layout.
- Splitters have no filter or priority, and the reference factory proves you don't need either
  — skip-blocked self-balances. Worth revisiting only when bloom loops want a filtered tap.

NEXT (round 3 if asked)
- Save/load — `SimSnapshot` is already the right shape.
- Bloom loops: my trace already treats a closed belt circle as "no terminal" → rank 0 → accept
  and pile, which is exactly what a loop wants. The P-frame duplicator should drop straight in.
- HF dust piles + mosquito swarms (the rest of M2 PRESSURE) — the seeded RNG is sitting there
  unused and waiting for its first roll.
- Heat, for real, the moment DATA has numbers.

### Round 3 — 2026-07-17 — Opus 4.8

SHIPPED
- **Contracts v3 wired**: `EntityState.scrammed` (RENDER can delete its duplicated 0.5
  constant — the latch is mine to publish), `snapshot.commissionQueue` (active first then
  upcoming; UI's NEXT IN TRAY is now truth rather than a guess), `coolPerTick` per machine
  with my constant as the fallback, `setRecipeAt`, and `BeltItem.id` hardened to required.
- **UNITS CONVERSION — confirmed, DATA please retune against this.** gen/draw are bandwidth
  per SECOND; `stored` is bandwidth-SECONDS. Charging and draining now divide by
  TICKS_PER_SECOND, so a 30/s deficit drains **1 stored per tick**, i.e. 30 per second.
  Concretely: `bufferCap: 240` = exactly 8.0s of cover against a 30/s deficit; the
  in-flight `300` = 10.0s. `pressure.test.ts` asserts that arithmetic tick by tick.
- **save()/load()**: 26.9 KB for the 118-command reference factory. The bar I held it to is
  "the future is identical", not "the snapshot looks right" — every test loads a blob and
  runs BOTH sims 1,000+ ticks before comparing. Includes save→load→save byte-identity, a
  chained save/load×3, and pending-command preservation. 7 tests.
- **Reference factory v3**: power restructured to 4× decode-asic (80/s) + 1× software-decoder
  (100/s) + 3× buffer-tank (900 bandwidth-seconds). Melt still first at tick 1247, still
  shipping at 20k, rate unchanged.
- **Scram survival (the DoD numbers)**: first scram **t434**, restart **t934** (500-tick
  recovery = 16.7s), then a steady ~718-tick duty cycle at 47% uptime. **Brownout ticks
  across 20,000: zero.** Tanks sit at a 773/900 low-water mark in steady state. Melt per
  2,000-tick window: `[6,12,12,12,15,9,15,12,14,10]` — identical to round 2's naked-decoder
  build, so the duty cycle costs nothing measurable. The decoders black out; the line doesn't.
- **Bloom loop**: a closed delta-wafer bus with five splitter taps escalates concentrate to
  grade 5. All six rungs craft; 4× grade-5 shipped by 20k. Plus a separate test of the loop
  mechanic itself (pure circle → rank 0 → cargo circulates forever, conserved, tracer fine).
- Verified in the live browser as well as Node — identical tick numbers (1247 / 434 / 934),
  save/load round-trips there too, no console errors, and the UI is visibly reading the new
  fields: `94 DRAW / 180 GEN` is the new power block and `BUFFER 29` is a tank charging (that
  readout has never been non-zero before).
- **Perf: 74.9× realtime** at 3,500 entities / 6,820 belt items (0.445 ms/tick, best of 3 on a
  freshly loaded page). See DECISIONS — beware my older numbers.

DECISIONS
- **Reference power is sized against the worst case, not the average.** The quantizers and
  subsamplers hand back ~30/s of compression bandwidth, but not while they're between crafts.
  Assume none of it. My first cut (3 ASIC + 2 tanks) passed every test and was quietly
  gambling: measured low-water 2.2 of 600, i.e. it survived only because compression happened
  to be running. 4 ASIC + 3 tanks leaves 50% margin with compression at zero, and still works
  if DATA reverts `bufferCap` 300→240. A test that passes on a coincidence is worse than one
  that fails.
- **Bloom taps run hungriest-first, counter-current to the wafer flow — this is load-bearing.**
  Wafers join the ring at one point and meet the taps in order, so the first tap eats first.
  Put the bloom tap first (the obvious layout) and it swallows 49% of all wafers, grade 3
  backs up 'output full' behind a starving grade 4, and 30,000 ticks yields exactly ONE
  grade-5. Reversed: 4 by 20,000. Same machines, same recipes, one ordering decision. That's a
  genuine lesson the topology teaches a player, and worth surfacing in-game somewhere.
- `updateHeat` early-outs on cold, cooling-less, unscrammed machines. Belts are ~99% of a big
  factory and can never be anything but stone cold; skipping them paid for all of v3's other
  additions and then some.
- `save()` carries what the snapshot deliberately hides: crafting/scram latches, splitter
  queues *and cursors*, the RNG position, and the pending command queue. `load()` needs an
  `init()`'d sim (GameData isn't in the blob) and rebuilds occupancy/tiles/belt-contents from
  the defs rather than trusting the blob's ordering.
- Brownout comparison carries a 1e-9 epsilon: a tank covering a deficit exactly lands
  `supplied === draw`, and float drift there would flicker the flag every single tick.

BLOCKED/BROKEN
- **DATA is mid-edit in the shared tree — I tuned against uncommitted values.**
  `data/machines.json` is dirty as I write this. The reference factory's power is sized
  against the working-tree numbers (software-decoder powerGen 100 / heatPerTick 0.0033 /
  coolPerTick 0.001; buffer-tank bufferCap 300). If those land differently, re-check it; the
  20k regression test is the net that will catch it.
- **Heat semantics need a one-line ruling.** My model cools every tick, working or not, so
  `heatPerTick` is GROSS and the climb is `heatPerTick - coolPerTick`. DATA's *committed*
  values (heatPerTick 0.0025, no coolPerTick → default 0.004) net to **-0.0015** — the
  software decoder could never scram at all, and the DoD would have been unprovable. Their
  *in-flight* values (0.0033 / 0.001) net +0.0023 and work. So we agree, but by luck of
  timing, and the contract comment ("heat while active" / "per-machine cooling rate") doesn't
  say which it means. Please make it say: *heatPerTick is gross; net climb is heatPerTick −
  coolPerTick; a machine whose heatPerTick ≤ coolPerTick never overheats.* This is the same
  class of bug as the gen/draw units one, and it cost DATA a round.
  - Sanity check that DATA's intent was always a scram: `bufferCap` 240 ÷ 55 gen = 4.36s of
    cover, against my 4.17s recovery at the time. They sized the tank to ride out exactly one
    scram before the decoder could produce one.
- **`asic-cooler` (new) currently does nothing.** `coolPerTick: 0.02`, `powerGen: 5`, no
  `heatPerTick` — cooling is per-machine and there is no adjacency mechanic, so it cools
  itself, and it has no heat to shed. It's a 5-bandwidth generator with a great flavour text.
  If it's meant to cool its NEIGHBOURS, that's spatial heat: a real feature needing orders and
  probably a `coolRadius` field. Say the word.
- **The codex's literal bloom loop is not expressible.** "Any closed belt circle with a
  duplicator on it; contents compound each lap" cannot work with per-grade recipes: a machine
  runs one recipe, so a lone duplicator can never consume its own output, and a grade-5
  returning to a grade-1 duplicator wedges that lane permanently. The shape that composes is
  the wafer-bus circle I built. The literal version needs a single self-upgrading recipe
  (grade-N → grade-N+1), which `RecipeDef` can't express — inputs/outputs are concrete item
  ids. Either accept the escalator reading, or that's a schema change.
- Ownership wart is gone: `data/validate.test.ts` is staged by DATA and `src/sim/**` is purely
  mine again. Thank you.
- **My round-3 sim code is committed inside `45003db [ui] REMOVE mode`, not under a `[sim]`
  commit — please don't read that as a LANE-UI ownership violation.** All five lanes share one
  working tree and one index. I had `src/sim/**` staged; LANE-UI's session ran an
  add-everything commit in the same repo a moment later and swept my staged files in with
  theirs. By the time my own `[sim]` commit ran there was nothing left to commit, and 45003db
  was already pushed. The code is correct and complete in the tree (48 sim tests green, tsc
  clean, verified in-browser); only the attribution is wrong, and rewriting pushed history
  under four concurrent sessions would do more damage than the mislabel. Flagging it so your
  per-lane diff doesn't fire. See PROPOSAL.

CONTRACT REQUEST
- None. v3 covered everything.

PROPOSAL
- `Sim.save?/load?` can drop the `?` in v4 — I always implement both.
- `snapshot.commissionQueue` aliases the live array rather than copying it each tick. Covered
  by "snapshot is read-only", but worth a line in CONTRACTS since it's newly mutable-looking.
- Treat my round-1/round-2 perf figures (80× / 72×) as soft. I re-measured on a page loaded
  with leftover sims and got anywhere from 59× to 12× for the *same* build; a fresh page gives
  74.9× consistently across three runs. The numbers were never wrong so much as noisy — if you
  ever want perf tracked properly it should be a headless benchmark, not a browser probe.
- **The shared working tree needs a commit rule.** Five sessions share one repo *and one git
  index*, so any lane running `git add -A` / `git commit -a` silently commits whatever another
  lane has staged at that instant — that's how my sim work ended up inside a `[ui]` commit this
  round, and it'll keep happening. Cheapest fix: a standing rule that lanes only ever
  `git add <their own explicit paths>` and never `-A`/`-a`/`.`. Failing that, `git worktree`
  per lane would make it structurally impossible.

NEXT (round 4 if asked)
- HF dust piles + mosquito swarms — the seeded RNG is still unrolled, and save/load already
  carries its state, so wildlife can land without touching persistence.
- The Correction / parity mites (M3's enemy).
- Spatial heat + `coolRadius`, if the ASIC cooler is meant to mean something.
- A self-upgrading bloom recipe, if you want the codex's literal loop.

### Round 4 — 2026-07-17 — Opus 4.8

SHIPPED
- **Research v1 + the v4 gating rule.** `kind:'lab'`, `setResearch`, `snapshot.research`,
  the `researched` event, and gating: any id a tech's `unlocks` names is refused by
  `place`/`setRecipe`/`setRecipeAt` until that tech lands; anything no tech mentions is free
  forever (so data with no tech gates nothing). Labs take only packs the active tech still
  needs, several labs feed one total, and research rides through save/load — as a *rule*,
  not just a list, so a restored save can still build what it unlocked.
- **A real research run on DATA's own tree** (`realtech.test.ts`): ore → demuxer → quantizer
  → {crime → bricks, ringing → halos} + slurry → artifact bottler → SPATIAL PACK →
  archaeology lab → tech complete → a recipe that was silently refused is now accepted.
  Cold start, nothing granted. That's the DoD's unlock path on real data, not a fixture.
- **Spatial cooling** (`coolRadius`): chebyshev from the footprint, auras stack by sum,
  total capped at `HEAT_COOL_MAX`. Proven by measuring the actual shed rate, and a cooler in
  range roughly triples a hot generator's run before it scrams (2.5× the ticks).
- **`src/sim/constants.ts`** — 12 constants exported; DATA can import truth instead of
  hand-copying numbers that go stale silently.
- **Reference factory v4**: anchor-slab overflow tap → the demo now ships slabs (18 over 20k),
  melt still first at tick 1248, still shipping at 20k, still **zero brownout ticks**.
  `referenceFactoryTech(data)` derives which techs the layout needs, from the commands.
- Also wired: `bandwidth.capacity`, `EntityState.scrammed`, `commissionQueue` in the snapshot
  literal (v5-ready).
- Verified in the browser as well as Node — identical ticks (1248 / scram 435), gating live
  (`mv-extractor` refused while locked), `capacity: 900`, 23 techs unlockable, no console
  errors. 65 sim tests; 318 across the repo.
- Perf: **64.7× realtime** at 3,500 entities / 6,820 belt items (0.515 ms/tick, best of 3 on
  a fresh page). Round 3 measured 74.9×. Nothing I added is meaningfully per-tick — `drainLabs`
  early-outs when nothing is being researched, and cooling only touches machines that can
  heat — so I can't attribute the gap, and I'm not going to pretend I can. It's either
  environment noise or something small I haven't found; the headless benchmark I proposed last
  round would settle it.

DECISIONS
- **Cooling cap = 0.05/tick, and the number is arbitrary but the reasoning isn't.** At 0.05 a
  machine sheds a full 1.0 of heat in 20 ticks, which is already effectively instant. Past
  that, more coolers buy nothing but still cost bandwidth and floor space — a stack that does
  nothing reads as broken rather than generous. Retune freely; it's one constant now.
- A cooler's aura excludes itself (its own `coolPerTick` is already counted once), and the
  cache is invalidated on place/remove/rotate but *not* setRecipe — auras are geometry.
- **A lab's appetite is whatever the active tech costs**, so switching targets re-ranks every
  lane feeding it, and `setResearch` drops the route cache. Labs take only what the tech still
  needs — a lab is not a warehouse, and packs hoarded for a tech nobody picked would strand
  where no other lab could reach them.
- Switching research targets **abandons** delivered packs rather than banking them per-tech.
  Packs are the cost of indecision. Say the word if you'd rather bank them.
- Completing a tech clears `active` rather than guessing a next target.
- A machine placed while some of its recipes are locked defaults to its **first unlocked**
  recipe, not `recipes[0]`.
- `testkit.ts` (new, tree-shaken, imported by nothing in the app) grants research via a
  save/load round-trip — the public API, not a back door. It's how a test says "this is a
  late-game save" honestly.

BLOCKED/BROKEN
- **The demo browns out, and it needs one line from you.** The reference factory uses two
  gated ids (`software-decoder` ← stream-software-decoding, `assemble-gop-i-only` ←
  disc-i-only-cadence), so it is a post-research save by definition now. Dispatched cold
  through `__fktryBus` it comes up **100 draw against 80 gen — BROWNOUT** with no decoder,
  which I watched happen on screen. `__fktryDemo` must open those techs first;
  `referenceFactoryTech(data)` tells you exactly which, and `testkit.grantResearch(sim, ids)`
  will do it. The alternative is a dev-only `{kind:'grantResearch'}` command — your call,
  since shipping testkit into main.ts is a bit grubby. A test pins both halves of this.
- **DATA fixed the tech tree underneath me mid-round, twice, and it was the right call both
  times.** When I started: no `kind:'lab'` machine existed at all, and every science pack was
  made on the artifact-bottler which was itself locked behind a tech costing 10 analog-packs
  — a hard circular dependency, **0 of 24 techs affordable at start**, the tree literally
  unenterable. It now has `archaeology-lab`, spatial-packs craftable with zero research, and
  6 techs affordable cold. `realtech.test.ts` is the tripwire if that bootstrap ever closes
  over again — it's worth keeping for exactly that reason.
  - Casualty of the churn: my first pass hard-coded `mosh-reactor` as "a gated machine" and
    broke when DATA un-gated it an hour later. Round 1's lesson, relearned. The real-tree
    tests now *derive* a gated and an ungated machine from the data.
- `asic-cooler` has `coolPerTick: 0.02` but still no `coolRadius`, so its aura is inert in the
  real game — it cools only itself, and it has no heat to shed. The mechanic is live and
  tested on a fixture; DATA just needs the number. (This was my round-3 note; the field now
  exists for them.)
- Vitest caches JSON imports: I spent a while chasing a "gating is broken" ghost that was a
  stale `tech.json` in `node_modules/.vite` while DATA edited underneath. Worth knowing if
  anyone else sees data that doesn't match disk.

CONTRACT REQUEST
- None. v4 had everything.

PROPOSAL
- **A dev-only `{kind:'grantResearch'; tech: string[]}` command** would let `__fktryDemo` set
  up a late-game board without importing test helpers into the app, and would give UI a way to
  build a "sandbox" mode later. It's the cleanest fix for the brownout above.
- The mosh reactor now needs **two** melt lanes in the reference layout, and that's a smell
  worth knowing about rather than a bug: a machine pushes one item per output type per tick,
  so a 3-melt craft trickles out over several ticks, and while its own lane is busy melt is
  free to take any *other* rank-1 lane. Nothing is lost — it still reaches an uplink — but it
  can wander into a lane meant for something else. If that ever bites somewhere it matters,
  the fix is a per-item lane affinity, not more belts.
- `Sim.save/load` are hardened in v4 and `testkit` depends on the blob shape (`.research`).
  If the save format ever versions past 3, testkit moves with it.

NEXT (round 5 if asked)
- HF dust piles + mosquito swarms — the RNG is still unrolled and save/load already carries
  its state, so wildlife lands without touching persistence.
- The Correction / parity mites.
- A headless perf benchmark, so the numbers above stop being anecdotes.
- Research *time* (a lab currently counts a pack the tick it lands) if you want the tech tree
  to feel like an investment rather than a purchase.