feat(tech-ablation): two-flag auto-apply path with 4-week reproduction gate (ROADMAP L2553)

[cipher813]

ROADMAP: L2553 — Tech weight ablation auto-apply — parallel-observation cutover (P1)

Cutover follow-on to alpha-engine-backtester#174 (tech weight ablation recommendation-only). The compute step is unchanged; this PR wires the gated write path that L2553 specifies.

Activation contract

Both flags default false → bit-identical behavior to today.

Flag (under tech_weight_ablation: in backtester config.yaml)	What fires	Live config
Both false	nothing	unchanged
use_tech_ablation_target=True (Stage 1)	shadow archive written every "ok" week to config/scoring_weights_per_sector_shadow_history/{run_id}.json + latest.json sidecar	unchanged
use_tech_ablation_target=True AND enforce_tech_ablation=True (Stage 2)	shadow archive + live write to config/scoring_weights_per_sector.json iff reproduction gate passes	written when same per_sector payload reproduces across last 4 shadow archives

Reproduction gate

_MIN_CONSECUTIVE_WEEKS = 4. Byte-equal per_sector match across the last 4 shadow archives required (the L2553 "4+ consecutive Saturdays" acceptance). One drift week breaks the streak — the gate explicitly does NOT tolerate intermittent shadow drift, matching the "recommendation must reproduce" framing.

Implementation

• optimizer/tech_weight_ablation.py:

  • init_config() + module-level _cfg (mirrors executor_optimizer).
  • _build_per_sector_payload() — recommendations: {team_id -> config_name} → {team_id -> {weight_name -> value}} via DEFAULT_GRID lookup. Unknown config_name drops cleanly (forward-compat guard).
  • _read_recent_shadow_archives() — lex-sorts YYMMDDHHMM keys descending; missing/corrupt archives skip with a WARNING (treated as "reproduction not yet reached", not a hard fail).
  • _check_reproduction_gate() — returns {passed, reason, n_consecutive}.
  • apply() — orchestrates the two stages.

• evaluate.py:

  • tech_weight_ablation.init_config(config) added alongside the other optimizers.
  • New _run_tech_weight_ablation() helper hangs compute + apply together (with --freeze short-circuit), replacing the prior inline lambda — mirrors _run_executor_opt / _run_weight_opt.

Tests (+12 new, suite 1694 → 1706 green)

• TestBuildPerSectorPayload × 3 — name→weights mapping including the unknown-name drop + empty-recommendations.
• TestApplyShadowGating × 4 — flag-off skips, status-not-ok skips, empty recommendations skips, shadow-only mode writes archive but NOT live.
• TestReproductionGate × 3 — insufficient history fails, 4-in-a-row matches pass, 1 drift breaks the streak.
• TestApplyLiveGating × 3 — enforce+insufficient history → shadow only; enforce+full reproduction → live written; enforce+drift in history → blocked + live untouched.
• One pre-existing test (test_recommendation_only_no_apply) reframed: the old apply_note string was the artifact this PR closes; new note points readers at apply() + the two flags.

S3 contract

Both keys are new additions — config/scoring_weights_per_sector.json and config/scoring_weights_per_sector_shadow_history/. No consumer reads them today, so this PR is forward-compatible on its own.

The research-side consumer (per-sector composite_weights override layer on top of scoring.yaml) is operator-owned: alpha-engine-research/config.py is gitignored per repo policy, and the per-sector schema is already in scoring.yaml (see _load_live_composite_weights_per_sector in this same module). The override-read can land when Brian flips use_tech_ablation_target=True and decides to wire the consumer.

Until then, the writes are pure observation — exactly the L2553 design intent for the recommendation-stability soak.

Data gate (not a code gate)

The compute step still emits insufficient_data until ~30 sub-score-populated rows/team accumulate; apply() correctly no-ops on that path. First non-empty recommendation is expected ~2026-06-27 per the research v15-migration accumulation timeline.

Composes with

• alpha-engine-backtester#174 (recommendation-only, this PR's prerequisite).
• alpha-engine-research v15 sub-score persistence (data gate).
• executor_optimizer.apply() (the pattern this PR mirrors).

🤖 Generated with Claude Code