Fix ExperimentMemory recording and surface silent failures as once-per-run degradations

[DermotOBrien-EC]

This is the focused resubmission invited in the #302 review: the ExperimentMemory constructor fix and the once-per-run degradation logging, fully decoupled from the result gate (which will follow separately, reworked per the review).

What this fixes

1. ExperimentMemory never recorded anything. execute_pipeline calls ExperimentMemory(store_dir=...), but the constructor is __init__(store, retriever, embed_fn=None), so every run raised a TypeError that the except Exception swallowed at debug level. The hook now builds a MemoryStore + MemoryRetriever and adds the ExperimentOutcome record type plus record_outcome() (persisted immediately).

2. Memory outcomes recorded metric 0.0 on every real run. The outcome hook extracted the metric from run_dir/results.json, a file production never writes (real per-run metrics live under stage-*/runs/*.json). It now aggregates the actual run artifacts via _collect_experiment_results and records the contract metric's mean, matching condition-prefixed keys such as ppo/primary_metric.

3. Silent failures become once-per-run degradation records. Experiment-memory init failures, outcome-recording failures, and knowledge-base export failures were swallowed (logger.debug or bare pass). Each now appends one {key, message} record to a degradations list written into pipeline_summary.json, with a single WARNING log the first time. Runs are never failed by these paths; the point is that the final summary tells the truth about what silently stopped working.

4. Degradations survive recursive PIVOT/REFINE runs. The rollback path re-enters execute_pipeline, and the outer frame's summary write would otherwise discard anything recorded inside the recursion. The accumulator is now threaded through the recursive call (a new optional kw-only parameter; all existing callers are unaffected).

5. Dead hooks removed. The EventLog and PitfallDetector hooks import researchclaw.pipeline.event_log and researchclaw.pipeline.pitfall_detector, neither of which exists in the repo, so both blocks were unreachable code inside try/except. Removed outright.

Testing

Every behavior above has a test, and each fix was verified red-first (the new test fails on the pre-fix code):

• memory init failure recorded exactly once
• outcome recorded with the real metric (0.42 from a stage-12/runs/run_01.json fixture matching the real artifact layout; the old phantom run_dir/results.json fixture is gone)
• outcome-recording failure recorded exactly once
• KB export failure recorded exactly once, pipeline completes
• a KB failure occurring only inside a recursive REFINE frame reaches the final pipeline_summary.json

Full suite: 2956 passed, 56 skipped (the skips are the env-gated live suites).