Correct errors found by review; hedge unmarked inferences

A four-agent review audited arithmetic, cross-document consistency,
epistemic honesty and style. It found substantially more than the
manual pass in the previous commit. Errors that changed conclusions,
not decimals:

Arithmetic
- production-timeline: the headline lead times did not reconcile with
  the stage table on the same page. Launch is 22-49 days (was 21-39),
  mature 11.5-22 (was 14-20), repress 6.5-11.5. Including gates,
  launch is 6-17 weeks, not 8-14.
- Consequence: at 17 weeks the launch configuration is at parity with
  EU/US plants, not ahead of them. The page now compares quote to
  quote rather than a stage sum to a quoted lead time, and says so.
- The galvanics turnaround saving is 1.5-5 weeks, not 3-5.
- scale.md Duplex capacity 300-600k/yr did not reproduce from its own
  stated basis; corrected to 175-540k. production-timeline disagreed
  with scale.md on the same fleet; both now agree.
- "300 units is under an hour of pressing" was out by an order of
  magnitude on the semi-auto line: 5-8 hours.
- project-timeline carried two totals for the same quantity. Now
  27-33 months throughout, with the Gantt's two ordering breaks fixed.
- scale.md: 300 kWp covers 28-45% of consumption, not 20-35%; roof PV
  stated as both 150-200 and 247 kWp for the same roof; parking canopy
  dimensions and PV disagreed with the RFQ.
- RFQ: fire tank stated 200 m3 where both its geometry (254 m3) and
  its own mass (250 t) say 250. Drainage now gives total sealed area
  66%, not the 46% paving figure, since that is what attenuation needs.
- Galvanics payback used only the top of the saving range: 1-5 years,
  not 1-4.

Inference marked as inference
- The 22,000 THB stamper figure is an invoice RESIDUAL. It carries
  lacquer cutting, test pressings, setup and margin, so it is an upper
  bound on the stamper component, not a measurement. Everything derived
  from it now says so, and a line-itemised invoice is the top question.
- The saving was also benchmarked against the wrong supplier: CNX buys
  from Stamper Discs, not a Thai reseller. Stated as 175-665 USD until
  that quote exists.
- The stamper-bottleneck hypothesis now names its alternatives. Jacket
  print runs would produce the same symptom and would constrain CNX
  equally, which would remove the competitive read entirely.
- Competitor capacity relabelled as nameplate times an assumed OEE,
  with the assumed shift pattern stated; all rows on one 300-day basis.
- The fixed/variable decomposition keeps the shape and drops the point
  estimates: the solve point inside the flat 1000-3000 tier is
  arbitrary and moves the fixed component by nearly half.
- "Nobody in Thailand plates" reduced to "we have found no evidence",
  with a desk check named as the action.
- Cut the USD 450-700k valuation of a competitor's presses entirely.
  It extrapolated a semi-auto quote onto a different machine, ignored
  the used market, and supported no decision.
- The collaboration payback divided capex by revenue where the same
  page uses contribution elsewhere: 5-10 years, not 1-3.

Missing caveats added
- No financing is committed or modelled anywhere in this repo;
  project-timeline previously asserted money was not a constraint.
- Leasing cannot pull the launch forward on its own, since the press
  chain governs and construction has float against it.
- The whole competitive analysis is one buyer, one plant, one
  afternoon, and the primary evidence is not archived in the repo.
- ResurRec is a competitor and may decline to buy from a new entrant.

docs/scale.md

@@ -10,7 +10,7 @@ page as real quotes and the site decision land.
 | Phase | Fleet | Rough output |
 |---|---|---|
 | Launch | 1 semi-auto press (Newbilt Single), 1 shift | ~90,000 to 135,000 records/yr |
-| Duplex | 2 semi-auto presses, 1 to 2 shifts | ~300,000 to 600,000 records/yr |
+| Duplex | 2 semi-auto presses, 1 to 2 shifts | ~175,000 to 540,000 records/yr |
 | Mature | 2 semi-auto + 2 automated (50/50 fleet), 2 shifts | ~1.0 to 1.5 million records/yr |
 
 Basis: the automated AD12 presses ~140 records/hour each (vendor datasheet);
@@ -20,16 +20,27 @@ realistic overall effectiveness of 60 to 70% applied. The theoretical
 ceiling of the mature fleet at 100% uptime is roughly 1.9 to 2.1 million
 records per year.
 
-**The 60–70% effectiveness assumption is worth challenging.** Newbilt
-publishes **≥95% yield and ≥95% uptime** for the Classic 12" automatic at
-a 28 s cycle (~128 records/h), and quotes ~30 s average for the Duplex.
-Even discounting the vendor's optimism and multiplying availability by
-yield by a realistic performance factor, that implies an OEE nearer 80%
-than 65% — which would make the capacity table here conservative by
-roughly a fifth to a third. Ask both press vendors for the **basis** of
-their uptime and yield figures (measured over what period, at what run
-length, counting which stoppages) as part of the re-quote. It is a free
-question and it moves the whole capacity model.
+**The 60–70% effectiveness assumption is worth challenging, but not yet
+changing.** Newbilt publishes ≥95% yield and ≥95% uptime for the Classic
+12" automatic at a 28 s cycle (~128 records/h), and quotes ~30 s average
+for the Duplex. Three reasons not to adjust anything on that basis:
+
+- **Machine uptime is not plant OEE.** A vendor uptime figure excludes
+  stamper and mould changes, label reels, colour changeovers, PVC lot
+  changes, first-article QC and breaks — precisely the losses that
+  dominate a short-run boutique plant. They are different quantities.
+- It would apply only to the **automated half** of the fleet. The
+  60–80/h semi-auto figure already embeds operator handling, so applying
+  an uplift to that half would double-count.
+- 0.95 × 0.95 is 0.90 before any performance factor; reaching 80% OEE
+  needs a performance factor near 0.89, which nobody has estimated.
+
+Ask both press vendors for the **basis** of their uptime and yield
+figures — measured over what period, at what run length, counting which
+stoppages — as part of the re-quote. It is a free question. If the
+figures survive scrutiny they would argue for a higher OEE on the
+automated fleet only, and the same logic would apply to any competitor
+running the same machine.
 
 ## Building
 
@@ -71,8 +82,10 @@ production, so either the main building shrinks by ~126 m² or the program
 grows into it. Decide which before the building is priced.
 
 Staff parking is roofed, in the front yard clear of the loading route:
-**10 car bays** (25 × 6 m canopy) and **~30 scooter spaces** (14 × 5 m),
-carrying a further **~32 kWp** of PV on their canopies. At 4.1 m and 3.5 m
+**10 car bays** (27.7 × 7.0 m canopy, 195 m²) and **~30 scooter spaces**
+(13.7 × 5.5 m, 76 m²), carrying a further **~40 kWp** of PV on their
+canopies — dimensions per the modelled layout in the
+[construction RFQ](finance/phase1-construction-rfq.md). At 4.1 m and 3.5 m
 they sit below the main building's 7.0 m low eave, so they cannot shade it.
 
 ### Expansion headroom: a second building
@@ -147,15 +160,17 @@ Grid connection: a medium-voltage connection with a transformer in the
 
 ## Solar and storage
 
-- Roof of a ~1,500 m² building fits roughly **150 to 200 kWp**; reaching the
-  illustrative 300 kWp in the utilities CSV needs carport or ground-mount
-  on the plot (about 1 rai of ground per additional 100 kWp, estimate)
+- The phase 1 roof carries **~247 kWp** (425 modules on the 1,523 m²
+  roof plane, per the [construction RFQ](finance/phase1-construction-rfq.md)).
+  This supersedes the older rule of thumb of 150 to 200 kWp on a
+  ~1,500 m² building, which the massing model has overtaken
 - Northern Thailand yield ~1,400 to 1,500 kWh/kWp/yr, so 300 kWp produces
   roughly **420 to 450 MWh/yr**
 - Plant consumption at mature two-shift operation: roughly **1,000 to
   1,500 MWh/yr** without electric steam, up to **~2,500 MWh/yr** with it
-  (estimate), so 300 kWp covers on the order of **20 to 35%** of
-  consumption; a bigger array shifts that up
+  (estimate), so 300 kWp covers roughly **28 to 45%** of consumption
+  without electric steam, falling to **~17%** with it; a bigger array
+  shifts that up
 - Battery: **271 kWh** Gobel stack (peak shaving and ride-through, roughly
   two to three hours of the chiller's electrical draw)