fix(learning): preserve exact candidates and research evidence

[drewstone]

Learning could reuse results after an experiment changed, accept malformed observations, and lose the errors needed for later learning.
This change binds results to exact inputs and preserves failures through code improvement, persistent lessons, and Knowledge research.

• Bind strategy checkpoints to profiles, tasks, settings, external execution identity, and authored source bytes.
• Create reflective proposal sources from the active checkout, signal, and cost account; reject invalid or partial patch batches.
• Preserve Git status columns, validate observations, expose storage failures, and serialize immutable Corpus writes across processes.
• Run the supplied Knowledge profile unchanged and align with released Eval 0.174.0 and Knowledge 14.0.0.
• Replace obsolete transfer-only architecture verdicts with separate specialist, domain-learning, evaluation-quality, and meta-learning claims.

Runtime 0.195.0 requires checkpoint.executionRef for strategy resume and createImprovementProposalSource(context) for reflective generators.
Bench 0.8.32 carries the matching dependency ranges.
The existing methods and package boundaries remain; this repair does not introduce another learning API or establish a capability gain.

Validation:

• Prior full Runtime integration: 3,561 passed, nine skipped on Node 24.11.1 with four workers.
  Later main integrations change instructions and skills only.
• Fresh typechecks, 701-file lint, package export/install/edge execution checks, version checks, release workflow checks, and documentation freshness passed.
• Clean exact-archive check passed 128 public imports and measured proposal creation with Sandbox 0.36.4 and 0.37.0.
• PrimeIntellect package integration passed.
• Built public-package smoke: eight malformed observations rejected, a valid empty response accepted, persistence failures retained, immutable Corpus values, and two exact Knowledge profiles preserved.
• Eval PR 738 and Knowledge PR 190 are merged and published.

The learning audit and retained findings distinguish implementation proof from unmeasured domain-learning and meta-learning claims.