v0.7.0

• Completed 7/7 architecture-class coverage: every stress-tested family now has
  a checked-in proof, anchored by the verified Granite-MoE-1B MLX port that
  closes the sparse-MoE class. The README, validation status, site, and
  llms.txt narrative were rewritten around this coverage and the four
  optimize-receipts/ observations, which record why the structured 8-bit pick
  held quality where the naive 4-bit default failed the gate. These remain local
  single-Mac, single-workload observations, not promoted benchmark claims.
• Landed the 2026-07-18 and 2026-07-20 nightly knowledge-curator runs as new
  dated research-run directories and retained bounded prior candidates instead
  of dropping them on refresh; the reviewed graph now holds 712 nodes and 505
  edges with the reconciled backlog and drift check still green.
• Added .github/workflows/deploy-site.yml: on a push to main touching
  site/ or VERSION, it deploys the static site to Cloudflare Pages with a
  pinned Wrangler and then polls the live data.js until it serves the released
  version, failing loudly if the deploy never goes live.
• Added .github/workflows/release.yml: when VERSION changes on main, it
  cuts the annotated vX.Y.Z tag and publishes a GitHub Release whose body is
  the matching CHANGELOG.md section, and no-ops when the tag already exists.
• Added .github/dependabot.yml for GitHub Actions and the .github/ CI pip
  requirements, both grouped, PR-limited, and gated by a 7-day supply-chain
  cooldown before a freshly published release can be proposed.
• Extended CI: a Python 3.13 lane joins 3.10 and 3.14 in the supported-Python
  matrix (with its hash-locked NumPy 2.5.1 wheel), and the release job now
  smoke-installs mlx-model-porting/requirements-tools.lock into a throwaway
  venv under --require-hashes --only-binary=:all: so the runtime lockfile is
  proven installable.
• Taught install_skill.py an antigravity client preset (the shared
  .agents/skills symlink root) and a --check mode that reports whether an
  installed skill's SKILL.md version matches VERSION without modifying
  anything, returning non-zero on a stale or missing install.
• Fixed a manifest time-of-check/time-of-use race: manifest.py now excludes
  the transient .converted-*.safetensors atomic-write staging files that
  convert_checkpoint.py creates and unlinks in place, so a check that races
  an in-progress conversion no longer tries to hash a file being written.
  Because these files are transient by construction and never distributed,
  skipping them cannot hide drift in a shipped file.
• Expanded the SKILL.md trigger map to route high-demand families directly:
  vision-language/multimodal, diffusion/flow, text-to-speech, ASR/streaming
  speech, sparse-MoE, and selective-SSM/hybrid inputs now each point at their
  runbooks, keeping the body under the audited size ceiling.
• Added tests/test_corpus_count_contract.py, which discovers the real suite
  size in-process and fails loud when README.md or VALIDATION.md advertise a
  stale offline-test count, and a repository conftest.py. Corrected the
  published counts across README, VALIDATION, and the research report to the
  discovered 551 offline tests and the current 712-node / 505-edge graph.

Deferred follow-ups

These deep-research gaps were identified during the release audit. Admitting
them to research_backlog.json requires new backlog_item graph nodes and a
graph regeneration (a new dated run), so they are tracked here for a future
release rather than hand-written into the reconciled backlog:

• an iOS/Swift on-device MLX deployment reference;
• a distributed / multi-Mac inference reference (mx.distributed);
• a runnable serving harness covering the mlx-lm server workflow;
• promotion of one stress-tested class (VLM or MoE) into the executable scaffold
  generator with an in-payload worked example; and
• a training/LoRA parity fixture, which today exists only as a contract.