v0.5.0 — signed verdicts, standalone verifier, verdict corpus

Added

• Signed verdict receipts (signalmap.receipt/1): every distill, fit and
  monitor run emits versioned JSON — claim, verdict
  (INCLUDED/EXCLUDED/PASS/REFUSED), evidence, input hashes, Ed25519 signature.
  The signing key lives in ~/.signalmap/signing_key (0600) and never enters
  the repo; only the public key travels in the receipt.

• tools/verify_receipt.py: standalone verifier that imports nothing from
  signalmap — stdlib plus cryptography. Checks signature, schema and internal
  consistency (a REFUSED verdict may not carry a deploy spec).

• Verdict corpus: the eight preregistered premium-family verdicts (rqa,
  coherence, envelope over six banks; 2 included, 6 honest exclusions) shipped
  as signed receipts in research/factory/receipts/, each pinning the sha256
  of the report it transcribes and labelled archive_signature — signed, not
  re-run. A suite gate fails if a shipped receipt goes stale.

• signalmap corpus rebuilds that corpus and prints the traction line.

• signalmap prove records the permutation-resolution floor of each split and
  reports inconclusive (cannot resolve) instead of null (no evidence) when
  a p could never have cleared the threshold at that group count.

• signalmap qualify (source profiling and method-family routing) and the
  compositional composer grammar (composer-v2).

• Published to PyPI: pip install signalmap. Releases upload through GitHub
  Actions via PyPI trusted publishing (OIDC), so no API token is stored
  anywhere.

• Test coverage for the seven product modules that had none — train,
  visualize, embed, the sinks, the store, discover, the simulator and the
  fit --dataset path. 116 -> 172 tests, coverage 78% -> 89%.

• Lint gate in CI (ruff, pinned version, explicitly selected rules).

• README section "Signed verdicts, verifiable without us": the receipt, the
  standalone verifier and the shipped corpus in three commands.

Fixed

• signalmap corpus printed the absolute path of every receipt, so a pasted
  listing carried the maintainer's home directory. Paths under the working
  directory now print relative to it.
• The CI job that proves standalone verification ran from the checkout root,
  where the source tree is importable through cwd: its "signalmap must not be
  importable" assertion fired on every run and the verifier was never
  exercised. It now runs outside the checkout.
• test_training_reduces_reconstruction_loss compared two independently
  initialised models and failed at random (1 in 3 parallel runs); both runs now
  start from the same seed.
• signalmap map crashed with IndexError on a single-recording dataset: the
  2D projection returned one column when the data could not span two. It now
  always returns two, and skips UMAP below three points.
• signalmap train --synthetic 0 reported a missing data source instead of
  naming the real problem, the frame count.
• causal.py linear-Granger design matrix used an out-of-scope loop variable
  after a rename; caught by the new lint gate.