0.2.0 — replay buffer, promotion gate, night scheduler, macros

Milestones 2 through 5, plus the inference layer that was split out of milestone 1.

Adapters trained with 0.1.0 are not comparable to adapters trained with 0.2.0: both the prompt formatting and
the default set of adapted modules changed. Retrain rather than carry old versions forward.

Added

• AdaptInference — AdaptSession: generation with or without the active adapter, digest verification before a
  load, adapter hot-swap that does not reload the model, cancellation mid-generation, and top-p and
  repetition-penalty sampling.
• AdaptData — SQLite replay buffer, privacy budget, TTL pruning with a durable prune log, and a scrubber
  pipeline (email, IBAN, card, phone) that runs before anything is stored.
• AdaptEval — the promotion gate: a held-out set pinned to the lineage, paired per-example cross-entropy, a
  one-sided Wilcoxon signed-rank test at α = 0.05, a minimum-evidence floor, and a broken-pin outcome that is
  neither a pass nor a failure. Abstaining and refusing are distinct in the type system, not two readings of one
  boolean.
• AdaptSchedule — AdaptPipeline (prune → sample → train → eval → promote), thermal, battery and memory
  policy, backoff on refusal only, and BGProcessingTask registration.
• AdaptMacros — @Personalizable(task:), @Prompt, @Completion.
• adapt-cli eval, and a script that seeds a seven-night registry by training seven adapters in seven separate
  processes, each resuming from the one before.
• Examples: StyleMirror (macOS demo) and QuickReply (iOS skeleton).

Changed

• Training and generation both go through the model's chat template. One formatter serves both paths, the
  convention is recorded in the adapter's metadata, and a session refuses an adapter trained under a different
  convention instead of generating subtly wrong output. On Qwen3-4B-4bit at 100 steps this moved the loss from
  9.63 to 1.49.
• LoRA target modules are explicit, and the default adapts attention projections only. Adapting the MLP
  projections as well costs 7.3M parameters against 2.6M for no visible gain in style, and reaches a lower training
  loss largely by memorising more. This changes the shape of a default-configuration adapter.
• EvalReport carries the gate's decision fields: the primary metric, its direction, and the example and
  supervised-token counts behind the number.
• Per-module error enums (AdaptCoreError, AdaptRegistryError, AdaptTrainError, …) instead of one shared enum.

Fixed

Twelve findings from an external review, the ones that could lose data or mislead:

• A checkpoint whose state file was not written last, so a crash mid-write could leave the registry unreadable.
• Gradient accumulation that weighted micro-batches equally regardless of how many tokens each contained.
• .noData returned when data merely ran short rather than out.
• LoRA initialisation that ignored the configured seed, making "same seed ⇒ same run" false.
• A missing digest verification before generate loads an adapter.
• Lineage IDs used in path construction without validation.

Known limitations

• On-device training has not been measured on a physical iPhone. The iOS build compiles and links and the pipeline
  runs in the simulator; step cost, thermal behaviour and the background window on real hardware are unmeasured.
• Encrypted sync between devices (M6) is not built.
• The demo app promotes through a provisional threshold, not AdaptEval's gate.
• The demo's blind test never completed in the app: launching straight onto the screen raced the state load, so
  generation never started. Fixed after 0.2.0 — see Unreleased.
• A rank-8 adapter over a corpus that is 20% Spanish and 20% Russian does not hold a non-English voice. The
  multilingual demo screen was cut rather than tuned.