Migrate specsmith → spec-kit + AEE; core on AEE library; GDELT Web Ngrams default

[tbitcs]

Summary

Three-stage migration, executed per the repo's governance (read-first:
AGENTS.md, docs/governance/rules.md, LIFECYCLE.md) and ledgered
per stage in LEDGER.md (append-only, AI-disclosed).

Stage 1 — specsmith → spec-kit + AEE (2e603341)

• specify init (v1.0.10, copilot integration) + aee & evaluator
  extensions installed from local clones as real files (verified: no
  gitlinks under .specify/).
• .specify/memory/constitution.md preserves Glossa's real governance:
  citation/provenance, append-only LEDGER, foundation-check gate,
  public/private correspondence boundary, AI disclosure, falsifiable
  claims, spec-driven flow.
• specs/001-glossa-lab-baseline/ documents the as-built platform
  (FastAPI + React, discovery engine, Evidence Graph, Indus status per
  README: 161 H+M readings, 90.96% coverage, preprint v4 DOI
  10.5281/zenodo.20414696).
• specsmith retired as active tooling (skills, scaffold.yml, tracked
  rate-limit state removed); runtime persistence .specsmith/ →
  .glossa-state/ in fetchers/base.py + model_intelligence.py.
  Historical mentions in LEDGER/CHANGELOG/ledger-archive untouched.

Stage 2 — core wired to the AEE library (a915eca5)

• Dependency: applied-epistemic-engineering>=1.0.4,<2.
• backend/glossa_lab/aee_core.py adapts extracted-claims JSON onto
  AEE Claim/Evidence/ClaimGraph + ScoringEngine.
  falsification_condition maps natively to
  Claim.falsification_tests; untested → AEE DRAFT (original kept
  in Claim.metadata); status enters scoring through attached evidence
  strength (the engine takes no status input — documented in the
  docstring).
• API, additive only: GET /api/v1/indus-evidence/claims/aee-scores
  and ?aee=true on GET /claims. Verified live: 31 claims, mean
  propagated score 0.485; default response shapes unchanged.

Stage 3 — GDELT guidance implemented (48c1ad8f)

• New default GDELT source gdelt_ngrams (Web Ngrams dataset):
  ~5-min-ago request, bounded 15-min-heartbeat walk-back (24 marks),
  quadgram matching with tri/bi/unigram reduction (>4-word keywords via
  constituent windows), topic exclusions, TOC cross-reference,
  .glossa-state/ watermark.
• DOC-API fetcher kept but opt-in only, docstring-noted as paused
  per GDELT's request during the Spanner migration.
• specs/002-gdelt-ngrams-and-frontier-methods/ also records (not
  implemented): a future fully-cited daily briefing over the discovery
  corpus, and manuscript-method design notes for future seal/tablet
  vision work (physical metadata up front; discrete focused passes).

Sync note

Branch includes a clean merge of origin/main after Dependabot
#50–#53 landed (workflow actions v7, Pillow bump) — no conflicts; AEE
dependency and Dependabot changes coexist.

Test results

• Full backend suite: 533 passed, 9 skipped, 0 failed
  (test_indus_evidence_api.py 25 passed run separately; the rest 508
  passed / 9 skipped).
• New tests: test_aee_core.py 18 passed, test_gdelt_ngrams.py 11
  passed (synthetic fixtures, no network). Ruff clean on changed files.
• backend/scripts/foundation_check.py cannot run off the Windows dev
  box (hardcoded C:\Users\trist\... path — pre-existing, untouched);
  noted in the stage-4 ledger entry.

Live smoke test (GDELT ngrams)

• Endpoint/naming verified against the dataset's documented example
  pair (2026-06-30 20:16 UTC): 1,063,647 quadgrams, 1,977 TOC entries;
  end-to-end matching on real data works ("disease" → 16 items;
  "Indus script" → 0 in that single minute).
• No current files exist: the 25-mark walk-back for today and spot
  marks over prior weeks all return 404 — the dataset appears not to be
  publishing at test time. The fetcher handles this correctly (bounded
  walk, no items, watermark untouched) and picks files up when
  publication resumes.

🤖 Executed by an AI agent (Muse Spark, via Muse) at the direction of
Tristen Pierson, per constitution §VI. Not merged — awaiting review.