test mode: PYAUTO_TEST_MODE_SAMPLES core API (phase 2)

[Jammy2211]

Overview

Phase 2 (of 4) of the test-mode size-realistic-samples feature: implement the PYAUTO_TEST_MODE_SAMPLES knob so PYAUTO_TEST_MODE=2/3 bypass runs write a samples.csv whose row count and byte size match a production Nautilus SLaM stage (N ~ 10k–100k), making cold SLaM runs instant while keeping resume-overhead profiling honest. The design is locked — implement D1–D4 from the phase 1 issue verbatim: #1378 (comment) (phase 1 record: PyAutoMind complete/2026/07/test-mode-representative-samples-phase-1-design.md).

Blocked at creation: PyAutoFit is claimed by aggregator-sqlite (#1377). Parked in PyAutoMind planned.md; starts when the claim frees.

Plan

• PyAutoConf (PR 1, merges first): test_mode_samples() accessor in autoconf/test_mode.py next to test_mode_level() — env var PYAUTO_TEST_MODE_SAMPLES, default 4, N < 4 raises ValueError.
• PyAutoFit (PR 2): re-export via the autofit/non_linear/test_mode.py shim; extend _build_fake_samples — N == 4 keeps today's literal branch byte-identical; N > 4 vectorized numpy synthesis (seed 0, prior-median row 0, logL_i = best − i, weights ∝ exp(−i/(N/10)) normalized) through the production Sample.from_lists → SamplesPDF → write_table path.
• Tests: the 7-item numpy-only list from D3 (exact 4-sample back-compat; summary/chaining; save→load round-trip; aggregator + database at N=1000; functional N=50k, no hard timing asserts; N=3 → ValueError).

Detailed implementation plan

Affected Repositories

• PyAutoFit (primary)
• PyAutoConf (accessor; merge first — dependency direction autofit → autoconf)

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoFit	main	clean — claimed by aggregator-sqlite (#1377)
./PyAutoConf	main	clean

Branch: feature/test-mode-representative-samples · Worktree: ~/Code/PyAutoLabs-wt/test-mode-representative-samples/ (created by /start_library once unblocked)

Implementation Steps

1. PyAutoConf autoconf/test_mode.py: add test_mode_samples() (D1 semantics; docstring in the module's existing register).
2. PyAutoFit autofit/non_linear/test_mode.py: add to import + __all__.
3. PyAutoFit autofit/non_linear/search/abstract_search.py _build_fake_samples: keep the N == 4 literal branch untouched; add the N > 4 vectorized branch per D2 (numpy (N, D) scatter p·(1+1e-3·G), zero-handling as today, best-first monotone logL, normalized exp-decay weights, log_prior = 0).
4. Tests per D3 in test_autofit/non_linear/ (numpy-only; no JAX anywhere).
5. Note in the PR's API-change summary: new env var, default-unset behaviour unchanged — phases 3 (autolens_profiling recipe) and 4 (docs) consume it.

Key Files

• PyAutoConf/autoconf/test_mode.py — new accessor (D1)
• PyAutoFit/autofit/non_linear/test_mode.py — shim re-export
• PyAutoFit/autofit/non_linear/search/abstract_search.py — _build_fake_samples (D2)

Guard-rails from the design (do not re-derive)

• Bypass write path applies no weight pruning (samples_above_weight_threshold_from is updater.py-only); the D2 weight floor w_min ≈ (10/N)e⁻¹⁰ ≥ 4.5e-10 keeps threshold-applying loads safe for N ≤ 1e5 (threshold 1.0e-10).
• Parity target measured: 10,187 rows × 21 cols = 9.07 MB (hst_fast source_lp[1]).
• Mode 3 (logL_best = -1e99) must work unchanged in both branches.

Original Prompt

Click to expand starting prompt

Test-mode representative outputs — Phase 2: core API

Type: feature
Target: autofit
Repos:

• PyAutoFit
• PyAutoConf
  Difficulty: medium
  Autonomy: supervised
  Priority: normal
  Status: formalised

Phase 2 of 4 of draft/feature/autofit/test_mode_representative_outputs_size_realistic.md.
Implements the D1–D4 decisions locked in phase 1 (..._phase_1_design.md) — do not
reopen them. Blocked until phase 1's decisions are posted.

Scope

• Implement the sample-count knob (per D1) — default 4, so existing behaviour is
  byte-identical when the knob is unset. Touch @PyAutoConf only if D1 placed the
  accessor next to test_mode_level() in autoconf/test_mode.py.
• Extend _build_fake_samples in @PyAutoFit
  (autofit/non_linear/search/abstract_search.py) to synthesize N samples per the
  D2 recipe. Vectorized numpy — no per-Sample python loops; N=50k must generate
  in seconds or the feature defeats its purpose.
• Unit tests for the D3 downstream contract at large N: samples.summary()
  quantiles, result.model / result.instance chaining, multi-batch structural
  guarantees; plus default-unset back-compat (exactly today's 4 samples). Library
  unit tests are numpy-only — no JAX.
• Aggregator scrape / database load on a large-N synthetic output verified as part
  of the ship gate (workspace_test if a library unit test can't cover it cleanly).

Ship

ship_library: PyAutoFit PR (+ PyAutoConf PR if touched, merged first per the
dependency graph). The PR's API-change summary unblocks phase 3.

[Jammy2211]

Library PRs created — implementation of the locked D1–D4 design is complete:

• feat: test_mode_samples() accessor for PYAUTO_TEST_MODE_SAMPLES PyAutoNerves#126 — test_mode_samples() accessor (merge first)
• feat: size-realistic test-mode bypass samples (PYAUTO_TEST_MODE_SAMPLES) #1381 — vectorized _build_fake_samples N>4 branch + 7 unit tests

Gate: PyAutoConf 138 passed; PyAutoFit full suite 1493 passed / 1 skipped; e2e bypass fit at N=10,000 in 1.7 s (samples.csv 10,000 rows / 1.31 MB, min weight 4.5e-8, chaining intact). Heart RED acked by human for PR-open (5 pre-existing unrelated reasons; same set as the 2026-07-16 CTI/COOLEST ships).

Workspace impact: added-only, default-off — no existing script uses the new env var, no migration needed. The downstream legs are already scoped as phase 3 (autolens_profiling pipeline_resume recipe) and phase 4 (docs), issued after these merge.

[Jammy2211]

Shipped and merged — PyAutoLabs/PyAutoNerves#126 (2026-07-17T06:29:36Z) then #1381 (06:29:44Z), per the dependency order. PYAUTO_TEST_MODE_SAMPLES is live on both mains. Next: phase 3 (autolens_profiling pipeline_resume recipe — the paused slam-resume-profiling task #70 is the natural consumer) and phase 4 (docs). Completion record: PyAutoMind complete/2026/07/test-mode-representative-samples-phase-2-core-api.md.