Deepen docs accuracy + repo-wide link guard

[blisspixel]

Second consistency pass, going deeper than the top-level 1.37.1 sweep. Auditing was done against source (config/models.py, cost_estimator.py, tier_registry.py, profiles.py, company_overview.yaml) via parallel agents; fixes applied and verified.

Durability

• tests/test_doc_links.py — every relative Markdown link in the repo (READMEs, guides, skills, design notes, examples), not just the doc index, must resolve or CI fails. Scanned 250 links; all resolve today, now permanently guarded.

Accuracy — scraping internals in INTERNALS.md had drifted from the code

• Tier order: was a 4-tier HTTP-first pseudocode; corrected to the real 9-tier browser-first escalation (tier_registry.py), which also matches ARCHITECTURE.md.
• Fingerprint profiles: was 4 profiles including a nonexistent Firefox and Chrome/122; corrected to the 5 real profiles (3 Chrome 130/131, Edge 131, Safari 18.2) with timezone/viewport noted as decoupled.
• Company Overview section count: 21 → 23 (added Industry Outlook, Board Perspective).
• Grading: was per-dimension "(0-25)"; corrected to a single 0-100 score.
• Strategies listing: refreshed to the full module set.

Accuracy — deprecated model IDs presented as current

• ARCHITECTURE XAI fallback row: retired grok-4.1 fast → grok-4.20-non-reasoning ($2/$6, the real GROK_MODEL_WRITING).
• EVAL judge example: retired grok-4-1-fast-reasoning → grok-4.3.

README polish

• Added Azure Foundry (AZURE_OPENAI_API_KEY) and AWS Bedrock (AWS_BEARER_TOKEN_BEDROCK) to the keys table + a primr keys test note.
• Surfaced the --acknowledge-host-agent-may-bill gate on the hybrid-pilot row.

Swept and confirmed clean

• Cost figures: fully consistent across all docs — no stray $5/$0.50 defaults, grounded-lite ~$0.30 everywhere.
• docs/design/*: audited; the dated decision records are accurate to 1.37.1 — left as-is.

Verified

• tests/test_docs_index.py + tests/test_doc_links.py pass; mkdocs build --strict clean.