feat(recall): ENGRAPHIS_RECALL_NARROW_ARM opt-in for lower-latency small-k recall

[Coding-Dev-Tools]

feat(recall): ENGRAPHIS_RECALL_NARROW_ARM opt-in

Rescued from ship/c5-narrow-arm-opt-in (abf92400) and rebased onto current main (post-#187/#189). 2 files, +221/−18: engraphis/core/recall.py + new tests/test_recall_narrow_arm_opt_in.py.

What it does

When the operator sets ENGRAPHIS_RECALL_NARROW_ARM to any non-empty, non-zero value, prompt-only recall for small-k callers (k ≤ 20) switches to a narrower first arm:

• first page: max(candidate_k, min(50, candidate_k * 2)) — ~2× less work at k=8 (16 vs 32)
• escalation ceiling: max(first_page, min(PROMPT_ONLY_MAX_CANDIDATES, candidate_k * 4)) — the second page can't silently undo the savings

The default wider arm (candidate_k + min(250, candidate_k * 3)) is untouched for everyone else, and the k ≤ 20 gate exists because the regression that motivated the widening only showed up on small-k callers; larger-k callers need full widening to find approved evidence.

Integration notes (conflict resolution vs #187/#189)

The commit conflicted with main's post-#189 arm-cap machinery. The combined block:

• honors the explicit operator cap (arm_candidate_k_cap= / env) in both narrow and wide paths — first page clamp and escalation-ceiling clamp
• preserves the test(e2e): make Galaxy sweep and clearance specs deterministic on busy runners #189 ceiling semantics (PROMPT_ONLY_MAX_CANDIDATES bound) in the wide path
• keeps reasoning_effort-style profile knobs and telemetry untouched

Verification

• tests/test_recall_narrow_arm_opt_in.py: 6 passed on this branch
• Full recall regression family green: test_recall (37), test_recall_arm_candidate_k_cap, test_planned_recall, test_recall_release, test_release_recall_contract
• ruff clean on touched files
• Default behavior unchanged unless the env var is set (fail-safe for the shipped 1.7 evidence)

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.