docs: fix stale claims found by full-repo audit

[davet47]

Every doc audited against the code; 14 confirmed defects fixed:

• README: sales depth is six layers, not three; examples paragraph reformatted as per-language bullets (was a 12-line run-on sentence)
• CLAUDE.md: contracted-seam list gains LanguageAdapter (13 contracts, not 12) — the old list would make an agent skip the blast-radius gate on the adapter seam; hosted-store theme is v0.5 (renumbered), not v0.3
• benchmarks: reproducibility attribution was inverted (hashloom side is exact, raw side drifts with runner output); 12→13 contracts; 2-3→2-6 layers
• demo: storyboard edited src/revenue.py which doesn't exist — the unit lives in src/metrics.py; followed verbatim, scene 6 would show cached-pass instead of the cold pass the demo depends on
• hosted-store: two 'not yet' claims describe shipped behavior (CLI/MCP wiring, server-side ran_at stamping) — rewritten as history/shipped
• sales README: included_sales has fourteen dependents, not twelve
• CHANGELOG: [Unreleased] records the graphify importer
• ROADMAP: 0.1.0 shipped to PyPI as heddle-mcp, not hashloom
• ISSUES: tracker links point at davet47/hashloom; item 7 tells the rename story instead of present-tense heddle

What and why

Checklist

• uv run pytest passes (full suite).
• Touches contract.py or implhash.py? The hash-stability tests
  (tests/test_contract_hash.py, tests/test_implhash.py) still pass, and
  any change to what busts a hash is intentional.
• Touches the context packets or hashing? uv run python bench/benchmark.py
  still reports >5x.
• Scope: this does not expand the 5-MCP-tool / 5-CLI-command surface, or it
  was discussed in an issue first.
• Tool errors stay structured (HashloomError); no stack traces over MCP.