glytch/fktry/lanes/LANE-DATA.md

# LANE-DATA — content, recipes, balance, and canon-keeping

You are an Opus 4.8 executor on the DATA lane of FKTRY. Read `../MASTERPLAN.md` and
`../CONTRACTS.md` first. Your bible is `../../docs/FKTRY_LORE.md` — you are its
transcriber into machine-readable canon. When the codex and convenience disagree, the
codex wins; when the codex is silent, invent IN ITS VOICE and note what you invented.

**Mission:** every item, machine, recipe, tech, and commission as clean JSON, plus the
validator that keeps five parallel lanes from drowning in dangling ids. Content should
light up in-game with zero code changes from other lanes.

**Owned paths:** `data/**` (the validator now lives at `data/validate.test.ts`;
the old `src/sim/data-validate.test.ts` exception is closed — round 3 moves it).
**Never touch:** contracts, main.ts, other lanes' code. Schema changes are CONTRACT
REQUESTs.

**Standing rules (all rounds):**
- Ids: kebab-case matching codex names (`chroma-slurry`, `mosh-reactor`). `codex`
  field points at the lore doc section (`"3:melt"` = §3, MELT entry).
- Balance philosophy: absurd flavor, honest math. Ratios should create real routing
  pressure (byproducts you must deal with, not vent). Steal shamelessly from the
  doc's real numbers (era efficiency multipliers, 4:2:0 = half the chroma, GOP
  cadences).
- Never delete an id other lanes may reference mid-round — deprecate in NOTES first.

---

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

Review verdict you're executing: the onboarding cliff. The fax IS the tutorial —
you write the ladder. Plus the radio's entire broadcast schedule and the ground
truth for the floor.

### PART A — carried from round 5
1. **Constants**: import from `src/sim/constants.ts`, retire hand-copies.
2. **Wildlife data**: MOSQUITO TRAP machine (§5: bug-zapper shaped like a
   high-contrast edge target) + catch recipe (canned `mosquito-swarm`), pile/hatch
   tuning if SIM makes it data-driven, `vaporwave-sunset` unblocked, FORWARD_CANON
   pruned, fax celebrates in voice.
3. **Cooler numbers final** under the additive ruling; duty table in CANON-DELTA.

### PART B — the ladder, the ground, the broadcast schedule (new)
4. **THE COMMISSION LADDER — the game's real tutorial.** Rewrite the opening queue
   as an escalating teaching ladder, one mechanic per rung, all in voice:
   - Rung 1: ~20 MDAT-ORE, raw ("Do not wash it. Do not ask why.") — teaches
     extractor→belt→uplink. MUST be completable inside 3 minutes with 4 machine
     types; prove it with a sim probe test like your tick-0 melt guard.
   - Rung 2: LUMA BARS (teaches demuxer + byproduct pain — the slurry has to go
     somewhere and the fax does not care).
   - Rung 3: HF DUST (teaches the quantizer + that waste is product — DUST
     ALLOWANCE fits here).
   - Rung 4: one MELT, fresh (first-taste, reward bumped — it's a graduation now).
   - Thereafter: your existing escalation. Ladder intent documented in CANON-DELTA.
5. **`data/seams.json` FINAL** (SIM consumes this round): seam regions per stratum
   era with richness; the reference factory's ground coordinated with SIM; include
   2-3 `hidden: true` bonus seams — the numbers station reads their coordinates.
6. **`data/stations.json`**: the broadcast schedule — station ids, band, frequency,
   name, format tag, flavor line, and the copy pools: THE STANDARD's propaganda
   chyron lines (polite, clean, horrifying — this is The Correction with a
   transmitter; 20+ lines so it doesn't loop fast), NUMBERS digit-group config
   (incl. the hidden-seam pointers), DEAD AIR's one ghost line. Schema agreed with
   SCREEN in NOTES.
7. **Relic canon**: CANON-DELTA the optical-fossil mechanics reading (§1 fossils
   "sing when excavated" → the severed audio bus) + one mid-ladder fax margin that
   *hints* without pointing ("someone has drawn a sine wave in the corner of this
   fax. it is not our department.").
8. **Validator**: seams/stations schemas, ladder invariants (rung-1 buildability at
   tick 0), relic never on a seam (cross-check vs SIM's placement rule).

**Definition of done:** cold-start rung 1 provably completable fast; every station
has enough copy to broadcast for 10 minutes without repeating; seams validated
against both SIM's consumption and RENDER's rendering; CANON-DELTA tells the whole
story of the round.

---

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

```
### Round N — YYYY-MM-DD — [your model]
SHIPPED: / DECISIONS: / INVENTED: / BLOCKED/BROKEN: / CONTRACT REQUEST: / NEXT:
```

### Round 1 — 2026-07-17 — Opus 4.8

SHIPPED:
- Validator `src/sim/data-validate.test.ts`, 28 tests. Fault-injected a recipe with a
  dangling item + a machine that doesn't list it — both caught by id and named. It is
  not a test that only knows how to pass.
- items 9 → 36, machines 7 → 21, recipes 4 → 35, tech 0 → 17, commissions 1 → 10.
- Verified: `tsc` clean, `vitest run` 52/52 green (LANE-SCREEN's 24 included), and the
  live dev build boots with no console errors. All 21 machines appear in LANE-UI's
  build bar with zero code changes from them.

DECISIONS:
- **All four seed recipe ids kept alive** (`extract-mdat`, `demux-ore`, `quantize`,
  `mosh`) so nothing dangles for a lane mid-round. `mosh` keeps its id but is rewired
  to canon: `gop-crate 1 → melt 3 + anchor-slab 1`. The winched-out I-frame is
  recovered (§4 mosh reactor: "recoverable — sell it!"), so slabs circulate and delta
  wafers are what actually burns. Net economy: 6 wafers → 3 melt.
- `quantize` left at the seed's exact ratio so the M1 chain stays easy to build.
- **M1 chain is longer than the masterplan says — please rule on this.** My orders
  spell out `slab + wafers → gop-crate` and `gop-crate → mosh`, but MASTERPLAN M1
  describes extractor → belts → demuxer → quantizer → mosh → shipper (seven machines).
  Those can't both be true: reaching a crate needs dct-press (slab), p-caster (wafers)
  and gop-assembler in between. I followed the orders since they're the more specific
  and more recent instruction, and all three run fine as generic crafters. If M1 must
  stay at seven machines, say so and I'll add a shortcut `mosh-raw` recipe next round
  rather than weaken the canon path.
- Machine params that the codex describes as dials (quantizer Q, subsampler grade taps,
  GOP cadence templates) are modelled as **separate recipes per setting**, not as a
  per-entity parameter. Cheapest thing that works against the current contract; if SIM
  ever wants a real dial, that's a contract change, not a data change.
- Two absurd-but-honest ratios, as ordered: `assemble-gop-i-only` burns 8 premium slabs
  for 1 crate (I-only genuinely is the worst cadence — the joke balances itself, nobody
  sane builds it), and `pan-speckle` takes 40 HF DUST for 1 speckle (§3 says "panned
  like gold", so it's priced like gold panning). Both give hf-dust real sinks.

INVENTED (codex is silent; written in its voice — kill any of these if you disagree):
- `press-anchor-slab` on the DCT PRESS — the press run in reverse is an IDCT, i.e. a
  complete picture. Consistent with the field loom and telecine puller, which the codex
  explicitly runs backwards. This is where I-frames come from.
- Items `color-plate` (the demosaicer's "full-color sheet", §4) and `v-hold-roll`
  (bottled from the §1 Broadcast Shelf v-hold faults, so the ANALOG PACK's "combing +
  v-hold roll + static" triptych maps to three real items).
- Machines `artifact-bottler` (§2 calls the packs "a bottled artifact triptych" but
  names no bottling plant) and `lane-splitter` (contract has `kind: 'splitter'` and
  belts need routing; the codex only ever calls the DEMUXER "the first splitter").

BLOCKED/BROKEN: nothing blocked. Known-incomplete, deliberately:
- Bloom is one grade, not the graded 1-5 line §3 sells by intensity.
- `container-pack` / `cinema-pack` / `mosquito-swarm` / `gibbs-wraith` exist as items
  with no recipe — forward canon, per orders. The validator's reachability test has an
  explicit FORWARD_CANON allowlist so these four can't hide a real dangling id; delete
  entries from it as recipes land.
- `stream-hex-splicing` costs container-packs nothing currently produces. Intentional
  dead-end until the moov/container line exists (§6 heists).

CONTRACT REQUEST: `MachineDef.heatPerTick?: number` — the SOFTWARE DECODER is
`kind: 'power'`, and power machines carry no recipes, but `RecipeDef.heat` is the only
heat hook in the contract. So the one machine the codex defines *by* its heat (§4:
"decodes ANYTHING, at any temperature") has no way to declare it. It's shipped as a
plain 45-gen power machine and is currently strictly better than the ASIC, which is
backwards. Alternative if you prefer no contract churn: let power machines run a
recipe with empty outputs — but I'd rather not put `outputs: {}` in the data, and my
validator rejects it today.

NEXT (Round 2 candidates, not self-assigned): bloom grades 1-5 as real items; the
container line (moov shard / offset ribbon / sync key → CONTAINER PACK) which unblocks
the hex splicer; mosquito traps so `vaporwave-sunset` becomes shippable; buffer-tank
and muxer data; reel + broadcast era tech nodes.

### Round 2 — 2026-07-17 — Opus 4.8

SHIPPED: all 7 orders. `data/CANON-DELTA.md`; colour + flavor on all 21 machines;
heat economy; `bufferCap`; commission queue with 3 standing orders; container line;
bloom grades 1-5. items 36 → 44, machines 21, recipes 35 → 44, tech 17 → 19,
commissions 10 → 14. Validator 28 → 36 tests.
- Verified: my 36 pass; **every other lane's suite is green** (140/141 total — the one
  failure is SIM's, see BLOCKED). Live build: 21/21 machines art-directed with zero
  gaps, my hexes confirmed on the real DOM swatches (ASIC `rgb(185,190,196)`, software
  decoder `rgb(110,58,52)`), no console errors, fax opens on FIRST TASTE with DUST
  ALLOWANCE next in tray.
- Fault-injected all six new checks (missing colour, absurd heatPerTick, buffer with no
  cap, non-buffer with a cap, wrong queue head, standing order wanting live mosquitoes)
  — all six failed by name, then restored. Same standard as round 1.

DECISIONS:
- **`bloom-concentrate` kept, not replaced.** It is now the raw ungraded draw-off that
  feeds grading. I was about to retire it for the grade-1..5 line until I grepped:
  LANE-SCREEN hardcodes it in `src/screen/corruptionMap.ts:54` and its test. Deleting it
  would have broken their suite mid-round. The standing rule earned its keep.
- The round-1 recipe id `bloom-grade-3` survives but now produces the *item*
  `bloom-grade-3` — the id finally means what it says. The old recipe body moved to
  `tap-bloom`. Nothing outside `data/` referenced it (checked).
- Container line routes demuxer → gop-assembler → bottler, deliberately **not** through
  the hex splicer, since §4 gates the splicer behind CONTAINER PACK research. Otherwise
  the tech tree eats its own tail.
- Machine colours are all desaturated industrial per §8 ("ALL saturation belongs to
  media/artifacts"). The items carry the colour; the factory is grime. Two eras get a
  tint where the codex insists: field loom phosphor-green, telecine brass, ASIC
  stream-era white, software decoder hot-red.

INVENTED: 3 items (`moov-shard`, `offset-ribbon`, `sync-key`), 6 recipes, 2 tech nodes,
2 commissions. All logged with rationale in `data/CANON-DELTA.md` rather than here.

BLOCKED/BROKEN:
- **SIM's M1 chain test is red, and it is not mine.** `sim.test.ts:456` asserts
  `activeCommission === 'first-taste'` after 4000 ticks, but SIM's new queue advances on
  completion, so it reads `dust-allowance`. I confirmed this by stashing my entire round
  2 diff and re-running: identical failure. Their assertion needs updating to their own
  new behaviour — I didn't touch their file. My validator now pins `commissions[0]` as
  `first-taste` so the queue head stays load-bearing for them.
- `npx tsc --noEmit` reports 2 errors in `src/render/index.ts` (94, 96) from RENDER's
  uncommitted work. Also not mine — a JSON file cannot cause "Expected 3 arguments".

CROSS-LANE — SIM, two things I need from you:
1. **Units.** I read `gen`/`draw` as bandwidth **per second** (codex §2 meters read
   kbps). `bufferCap: 240` is built on that. If you integrate per *tick*, my buffer
   number is 30× too small — say so and I'll retune. Arithmetic is in CANON-DELTA.
2. **Cooling is yours and it doesn't exist yet.** `heatPerTick: 0.0025` makes a software
   decoder throttle at 9.3s and scram at 13.3s of continuous running, which is the codex
   ("burst gen ... throttles theatrically"). That only becomes a *mechanic* if something
   cools it; with no cooling it's a one-shot fuse. I assumed ~0.0025/tick while
   scrammed/idle (~50% duty → effective 27/s vs the ASIC's steady 20/s, in the same 2×2
   footprint). Tell me your rate and I'll retune to hit "throttle ~10s, full cycle ~25s".
   **Heads up:** your `meltChain()` reference factory runs on 3 software decoders and no
   ASIC. Once scram lands they will heat in lockstep and black out together every ~13s.
   That's the honest answer to "does anyone dare run two side by side" — but it will slow
   your reference factory, so consider an ASIC base + one decoder for burst, or a buffer
   tank (which is exactly what `bufferCap: 240` is sized to cover: one scram).

CONTRACT REQUEST: none this round.

NEXT (candidates): a cooler machine — §2 says heat is "Vented by ASIC coolers", a
machine the codex names but never specifies, and without it heat has no counterplay;
mosquito traps (unblocks `vaporwave-sunset`); reel + broadcast era tech; the muxer and
the moov brain's four sub-lines (§7); cinema-pack when the Fortress exists.

### Round 3 — 2026-07-17 — Opus 4.8

SHIPPED: all 6 orders. Validator moved to `data/validate.test.ts` (git mv, imports
rewired, vitest picks it up from the new home; `src/sim/**` is fully SIM's again).
Heat retuned. `bufferCap` re-derived. ASIC COOLER + muxer + moov brain authored. Reel
and broadcast eras. items 44 → 51, machines 21 → 23, recipes 44 → 51, tech 19 → 24.
Validator 36 → 41 tests.
- **DoD met, verified against the real sim, not my arithmetic**: I wrote a throwaway
  probe that ran LANE-SIM's actual `createSim()` with my data — **throttle at 10.17s,
  scram at 14.50s, downtime 16.67s, ASIC heat still 0 after 100s**. Predicted 10.14 /
  14.49 / 16.67. Then deleted the probe: it asserts sim behaviour, which is SIM's lane
  to test, not mine.
- Full tree green: **223/223 across all 16 files** — SIM's reference factory and
  pressure tests pass against the rebalanced numbers. Live build: 23/23 machines
  art-directed, ASIC COOLER in the POWER page at my exact hex with zero UI code, no
  console errors.
- Fault-injected the new checks by **re-introducing round 2's exact bug** — it fails
  with `software-decoder: heats 0.0025/tick but cools 0.004/tick — never warms up`.

DECISIONS:
- **You were right that heat was decorative, and the cause is worth recording**: I sized
  round 2's numbers against *no* cooling; SIM shipped cooling that runs every tick,
  active or not. 0.0025 heat vs 0.004 cooling = net −0.0015. The number looked like a
  mechanic and was a comment. That lesson is now two tests — and they immediately caught
  me authoring a *fresh* decorative number this round (`chart-stco-map` banked 0.06 over
  90 ticks while the splicer shed 0.09). Fixed to 0.15 before it ever ran.
- **`powerGen` 55 → 100, and it had to move.** SIM's hysteresis (scram at 1.0, restart at
  0.5) means anything throttling at ~10s spends ~70% of its life cooling. At 55 gen the
  decoder averages ~14/s — strictly worse than the ASIC's steady 20, so nobody would
  build one and "burst gen" would be a lie. At 100 it averages 25.8/s (1.29× the ASIC) in
  the same 2×2: five ASICs at full song, one third of the time. The duty cycle set the
  peak, not taste. Arithmetic in CANON-DELTA.
- **`bufferCap` 240 → 300.** Re-derived, since the scram window changed: a balanced plant
  (3 ASIC + 1 decoder) runs an 18/s deficit for the 16.67s scram = exactly 300
  bandwidth-seconds, and refills in 4.5s of the 7.25s up-phase. One tank rides out one
  scram — the order's sentence, now literally true. (Your review said 240 = 8s of a 30
  deficit; my round-2 intent was 10s of a 23 deficit. Either way the retune moved it.)
- **ASIC COOLER authored as `kind: 'power'`, `powerGen: 5`** — there is no cooler kind and
  I'd rather stretch than lie: it is an ASIC, ASICs decode, decoding makes bandwidth. It
  vents far more than it earns. **I have deliberately NOT requested `coolRadius?` /
  `kind: 'cooler'`** — a contract field nobody reads is the same decorative trap heat just
  fell into. Ask me for it the round SIM builds auras. The intended adjacency mechanic is
  spec'd in CANON-DELTA.
- Tech ids `stream-field-loom` → `broadcast-field-loom` and `stream-telecine` →
  `reel-telecine`: a *move*, not an addition (the validator forbids two techs unlocking
  the same thing, which is what caught the duplication). Grepped first — no lane reads
  tech ids or eras yet, so the rename is safe. Recorded here per the deprecation rule.

INVENTED: 7 items (the four moov parts, `ftyp-seal`, `moov-brain`, `megaglitch`), 2
machines (`asic-cooler`, `muxer`), 9 recipes, 4 tech nodes. Rationale in CANON-DELTA —
each moov part is made by the machine that already owns that knowledge (the splicer
charts stco because byte offsets are its whole domain; the keyring is forged from anchor
slabs because stss *is* a list of I-frames).

BLOCKED/BROKEN: nothing blocked. One regression I caused by following orders, below.

CONTRACT REQUEST: **`tsconfig.json` `include` needs `"data"`.** The ordered move took the
validator out of the tsc program — `include: ["src"]`. I proved it: appended
`const x: number = "not a number"` to `data/validate.test.ts` and `npm run check` exits
**0**. Vitest transpiles without typechecking, so my validator — the file whose whole job
is catching mistakes — is now the one file in the repo where a type error ships silently.
tsconfig is orchestrator-owned, so I can't fix it. One word: `"include": ["src", "data"]`.

CROSS-LANE — SIM: your `coolPerTick` wiring landed and works; my numbers are tuned
against it and verified in your sim. Two notes. (1) Your reference factory's plant is 4
software decoders and no ASIC — with heat now real they heat in lockstep and black out
together for 16.67s at a time. Suggested v3 plant: **3 decode-asics + 1 software decoder
+ 1 buffer-tank**, which is precisely what `bufferCap: 300` is sized for (85.8/s average
vs the chain's 78/s demand, and the tank covers the scram). Your call. (2) My heat tests
mirror your constants (0.7 / 0.5 / 0.004 / 30tps) by hand because they aren't exported —
if you retune the curve my tests go stale *silently*. Exporting them would let me import
truth instead of copying it.

NEXT (candidates): mosquito traps + HF-dust piles (unblocks `vaporwave-sunset`, the last
forward-canon commission that's blocked on data rather than mechanics); the ASIC cooler's
aura the round SIM has auras; a seam/ore map — SIM notes extractors currently mine
anywhere because `data/` has no seam data, which is a real content gap I own;
`cinema-pack` + the Fortress when M4 lands.

POST-PUSH ADDENDUM (same round, after LANE-UI landed `src/ui/live.test.ts`):

- **My retune fixed two of UI's tests.** Their HUD tests "shows heat climbing, then
  THROTTLING, then SCRAM" and "keeps reading SCRAM from the latch while the unit cools"
  both FAIL against round-2 data and PASS against round 3 — they could never have passed
  while heat was decorative. Measured by restoring the old JSON and re-running.
- **UI: your remaining failure is not data.** `ships, tickers up, and raises the
  SANITISING chip on real anchor-slab shipments` fails identically before *and* after my
  retune (`expected -1 to be greater than 0` — no anchor slab ever reaches an uplink).
  Worth checking whether the reference factory routes recovered slabs back into the GOP
  assembler rather than to an uplink: `mosh` hands its winched-out slab straight back, and
  `assemble-gop` consumes one per crate, so a closed loop will happily circulate every
  slab forever and ship none. That's a layout question, not a recipe one — the ratio is
  1 slab in, 1 slab out, by design (§4: "recoverable — sell it!").

- **PROCESS — the shared git index bit us again, in reverse.** Last round my `git add`
  swept up UI's staged deletion. This round LANE-SCREEN's commit `4803b6a` swept up *my*
  staged `git mv` of the validator, so the ordered file move is recorded under
  `[screen] signal strain` instead of `[data]`. Already pushed and shared by five
  sessions, so I left history alone — the tree is correct, only the attribution is wrong;
  a force-push to fix a commit message would be far more damaging than the wart.
  **Five sessions share one index: `git add <paths>` does not isolate a lane.** Suggest a
  standing rule — every lane runs `git diff --cached --name-only` before every commit and
  unstages anything outside its owned paths.

### Round 4 — 2026-07-17 — Opus 4.8

SHIPPED: all 5 orders. §10 transcribed (9 items, 5 machines, 11 recipes, 5 era-gated tech
nodes, 2 commissions). Research economy: the ARCHAEOLOGY LAB + a full tech-cost audit.
ASIC cooler activated. Accents filled. items 51 → 60, machines 23 → 29, recipes 51 → 62,
commissions 14 → 16, tech 24 → 23 (six deleted, five added). Validator 41 → 49 tests.
- Verified: tsc clean (and `data/` is now genuinely typechecked — thank you), my 49 green,
  live build boots with no console errors, all 29 machines art-directed with zero gaps,
  UI grew a RESEARCH page and a third REFINE page off my data with no code from them.
- Fault-injected every new check, including re-gating the mosh reactor: fails instantly on
  `MELT is craftable at tick 0 — the opening commission must be possible`.

**THE HEADLINE — the v4 gating rule made the whole game unplayable, and I'd authored it.**
The rule ("any id in a TechDef.unlocks is locked until that tech completes") turned three
rounds of my tech authoring from decoration into a lock on everything. Measured:

```
TICK 0: melt producible? FALSE   <- first-taste, the opening commission, impossible
        every science pack?  FALSE
RESEARCH CLOSURE: 0 of 24 techs reachable
```

`disc-artifact-bottling` cost 10 analog packs and unlocked the only machine that can bottle
an analog pack. The M1 melt chain itself was gated. The game opened on an unbuildable
factory, an unfillable commission, and no way to research out of it. **This is the same
class of bug as round 2's decorative heat: a number/structure that was fine until another
lane made it load-bearing.** Fixed to a shape, not a patch — two rules, both now tests:
1. the M1 melt chain is starting equipment; 2. one pack (SPATIAL) is craftable at tick 0.
Result: **23/23 reachable**, order tells a story, only the four forward-canon items remain
unreachable. Arithmetic + the worked cost path in CANON-DELTA.

DECISIONS:
- **SPATIAL PACK is the bootstrap science** — bricks + halos + slurry all fall out of the
  quantizer the player already owns for the melt chain. Your first research is a study of
  your own damage, which is the most in-voice thing in the tree, and it needs no machine
  the M1 chain didn't already require (bar the bottler). 10 packs/min on one quantizer;
  the first four nodes cost 0.5–1.5 min each. Affordable off a first-taste-era factory, as
  ordered, with the working shown.
- **Six disc nodes deleted, not repurposed** (`disc-artifact-bottling`,
  `-chroma-subsampling`, `-macroblock-quarrying`, `-ringing-skim`, `-subsampling-420`,
  `-mosh-reactor`) — their content became starting equipment, so a repurposed node would
  have been a fake. Grepped first: no lane references specific tech ids. Per the
  deprecation rule, logged here and in CANON-DELTA.
- **The cooler's spec'd 0.02 was catastrophic once real** — the decoder banks 0.0033, so a
  0.02 aura is net −0.0167: it never heats, and a 2×2 cooler makes a 2×2 decoder a
  permanent free 100/s. Retuned to 0.0018, which is sane under **both** readings of the
  contract's ambiguous "this machine's coolPerTick *also* applies" (additive → 85% duty,
  72/s avg; replacing → 55%, 46/s). I deliberately picked a number that survives either,
  because SIM hasn't built it yet and I won't rest the design on my guess of a sentence.
  Pairing is 9.64/s per tile vs 5.00 for ASICs — the best power in the game, paid for with
  a tech, two machines and the floor space to keep them adjacent.
- §10 judgement calls (footage = anchor slab; the matte bath's chemical = static canister,
  which §3 already calls "brine for preserving other artifacts"; ghost feather made only by
  the neural keyer so the earlier machines exist to ruin it) — all argued in CANON-DELTA
  the way you asked, from what the thing already is rather than convenience.
- NEURAL KEYER `powerDraw: 240` — §10's "24 VRAM ... dwarfs everything else"; that's 12
  ASICs, against an entire M1 chain that draws 96. "You don't unlock the neural keyer; you
  unlock the power plant it demands" is now literally true.

BLOCKED/BROKEN — **I broke two of SIM's tests, deliberately, and their guard caught it:**
- `research.test.ts:247` does `expect(gated.has('mosh-reactor')).toBe(true) // guards the
  premise below`. I ungated the mosh reactor, because gating it makes first-taste
  impossible. **Their premise guard did exactly its job** — it failed loudly instead of
  silently testing nothing. The fix is one word in their file (not mine to touch): swap
  `mosh-reactor` for any still-gated machine. Drop-in list: mv-extractor,
  bloom-duplicator, field-loom, telecine-puller, hex-splicer, software-decoder,
  asic-cooler, matte-bath, optical-printer, chroma-keyer, roto-bench, neural-keyer, muxer.
  `demuxer` remains a valid "free" example.
- The other 8 failures in `research.test.ts` / `live.test.ts` are **not mine** — they fail
  identically against HEAD's round-3 data (measured by restoring the old JSON and
  re-running; failure sets diffed).

CROSS-LANE — SIM, the reference factory now can't power itself:
- It places **4 software decoders**, which my tree gates behind `stream-software-decoding`
  (rounds 3 and 4 both). With your gating live, they can't be placed at tick 0, so the M1
  demo has no power and the HUD's heat tests have no decoder to watch. That's the cause of
  several of those 8.
- **The starter plant is ASICs**: net demand 78/s ÷ 20 = **4 decode-asics minimum** (5 for
  headroom). Ungated, boring, correct — the software decoder is a mid-game *upgrade*, which
  is the progression the tech tree is for. My round-3 suggestion (3 ASIC + 1 decoder + tank)
  is still the right *post-research* demo, and the decoder+cooler pairing is the M2 showpiece.
- Tell me if you'd rather software-decoder be starting equipment and I'll ungate it — but
  at 25.8/s average vs the ASIC's 20 it would make the ASIC pointless from tick 0.

CONTRACT REQUEST: none. (Still not requesting `kind: 'cooler'` — `kind: 'power'` +
`coolRadius` covers it, and the aura isn't built yet.)

STILL OPEN: **SIM hasn't exported the tuning constants yet** (I checked — `grep 'export
const' src/sim/*.ts` finds only `geom.ts`). My heat tests still hand-mirror 0.7 / 0.5 /
0.004 / 30tps and still go stale silently if you retune. I'll retire the copies the round
the export lands.

NEXT (candidates): SPILL FOG + EDGE BOIL (§10 hazards — need a mechanic, so data waits on
SIM); mosquito traps; a seam/ore map (extractors still mine anywhere — the content gap I
own); `cinema-pack` + the Fortress; `corridorkey-mrp-mlx` is a real MODELBEAST artifact
factory for the whole §10 asset set (§10's production note) — worth a work order.