NATIVE mode and FAST-PATH atomic paths produce no LEARN reflection signal

[virtualian]

Observation

The Algorithm's LEARN phase (phase 7/7 in PAI/Algorithm/LATEST v3.7.0) is the sole writer of structured reflection entries to ${PAI_DIR}/MEMORY/LEARNING/REFLECTIONS/algorithm-reflections.jsonl. That JSONL is the input for MineReflections, AlgorithmUpgrade, and Upgrade workflows — the closed-loop learning mechanism.

Two execution paths bypass LEARN entirely and therefore contribute nothing to the learning signal:

1. NATIVE mode — per CLAUDE.md, NATIVE is selected for "single-step, quick tasks (under 2 minutes of work)". The Algorithm file is never loaded, so no phase runs and no reflection JSONL is appended.

2. FAST-PATH atomic classifier — inside ALGORITHM mode, the FAST-PATH table (LATEST line 196) explicitly says atomic actions "Skip LEARN reflection."

Why this may matter

• The learning loop only sees Standard+ non-atomic ALGORITHM runs.
• The most frequent category of real work — quick tasks and atomic operations — is invisible to reflection mining.
• Systematic patterns in short-task failure modes (wrong-target edits, scope creep on atomic asks, misread intent) are exactly the kind of signal the scope gate and wrong-target rules were added to address, but those failure events in NATIVE/FAST-PATH paths never reach the JSONL that drives rule evolution.

Questions

1. Is this intentional (cost/ceremony tradeoff) or a gap?
2. If a gap: should NATIVE and FAST-PATH paths emit a minimal reflection entry (e.g. task + effort=native/atomic + implied_sentiment) so the learning loop sees them?
3. Is there a separate signal source covering these paths that I've missed (ratings.jsonl, hook-emitted telemetry)?

Evidence

• PAI/Algorithm/LATEST lines 426–447 (LEARN phase + JSONL append)
• PAI/Algorithm/LATEST line 196 (FAST-PATH "Skip LEARN reflection")
• CLAUDE.md NATIVE mode definition (Algorithm not loaded)