ArWen 2.6.1

Continuous assimilation gets a door, agents get a protocol, and the
sub-kilometre unlock is finished.

New:

• gpuwm cycle: continuous radar assimilation as a door. A cycle clock
  schedules analyses, a ledger records every cycle's inputs and verdicts,
  a supervisor drives the fetch, assimilation and forecast legs, and
  anchored child domains follow the feature being cycled.
• An MCP server: arwen-mcp serves the CLI doors as tools over stdio, so
  an LLM agent can plan, price, fetch, prep, forecast and render. Every
  tool runs the real command line or reads a receipt a door already
  wrote; refusals arrive as the CLI's own sentences. Long work runs as
  detached jobs that survive a server restart, one GPU job at a time per
  card. pip install gpuwm[mcp].
• The MPAS column-batch seam carries P3: microphysics_scheme = "p3"
  transports the scheme's eight species through both phases beside the
  unchanged WSM6 default, and a restored P3 seam continues
  bit-identically.
• Sub-kilometre static builds: below a 1,310 m source-pixel diagonal the
  static builder now derives categorical fields by supersampling, keyed
  to each mesh's measured minimum edge, where it used to refuse.
  Demonstrated end to end with a 937.5 m mesh generated, built, and run.
• Level-2 regional spectral numerics: an optional [spectral_numerics]
  table runs scale-selective hyperdiffusion and divergent-mode damping in
  off (bitwise inert, the default), shadow (receipts only), and
  apply modes, with per-step receipts bound into the run capsule and a
  gpuwm spectral-op evidence door. A research instrument, not endorsed
  physics.
• Nowcast skill is scored against an operational baseline, with structure
  metrics beside the existing point verification.
• Windowed all-radar ingest: assimilation cost follows the sites in the
  active window, not the domain, and successive nest windows hand their
  interior over during the overlap instead of restarting cold.
• Nineteen measured sizing configurations for ensemble, reanalysis and
  AI-model sources the configuration set had not covered.

Fixed:

• gpuwm prep no longer grows its host memory with the length of the
  forcing window: the mapped engine decodes one valid time and drops it
  before asking for the next, and the identity pass admits objects under
  a bounded byte budget. An ICON-EU preparation peaks at 2.32 GiB at
  both a 3 h and a 6 h window, frames byte-identical; compressed
  field-per-file sources still pay a modest wall premium for the
  per-time discipline, and that remainder is named work.
• gpuwm check on a nested tree with [tiles] prices the mixed road
  the run door actually takes (child streamed, parent resident) instead
  of refusing trees that fit; the report names each domain's road and
  --json publishes it.
• A [tiles] configuration the loader refuses reaches the terminal as a
  refusal with its remedy intact, on every door, instead of a traceback.
• A restored MPAS column-batch seam no longer refuses its first
  radiation-not-due step: the restart payload carries radiation carrier
  provenance, and the seam's device round-trip gate runs at every cut.
• P3 no longer computes NaN supersaturation ratios on its first step:
  the cold-start saturation mixing ratios are floored, default-on in all
  three arms, inert from step 2 onward, with the one step-1 delta
  documented.
• mp=50 composes with RTE+RRTMGP radiation: the adapter couples P3's own
  two-radius cloud optics, matching the scheme's own inventory, and the
  refusal that outlived the gap is retired.
• A wrong-version P3 lookup table is refused as a version mismatch
  instead of being reported as byte corruption, and gpuwm doctor
  checks the table.
• Velocity dealiasing uses each radial's own Nyquist velocity; the
  reader detected per-radial disagreement and then dealiased against one
  scalar anyway.
• A radar with no innovations no longer reports the whole network's
  residual as its own, and the continuous nowcast daemon no longer
  crashes at the first observation stage of every cycle.
• Data assimilation reads P3's own 10 cm reflectivity instead of
  refusing an mp=50 background, and clear-air reflectivity carries one
  floor value instead of two.
• A cycling analysis takes the card by default when one is reachable
  (auto); host and cuda stay honoured verbatim and every receipt
  records which device ran and why.
• The P3 oracle suite now fails when physical processes are switched
  off; the previous fixture set stayed green with five disabled.
• Mesh generation receipts print small trace values in scientific
  notation, and the ladder's location check credits grading and
  insertion instead of refusing meshes it graded correctly seconds
  earlier.