[Jammy2211]

Summary

Phase 2 (workspace record) of interferometer-numba-route-gate. The Aris routing table on both masks, re-measured under PyAutoArray#632's default cost model (PyAutoArray main 17caf228), plus the campaign's owed numba-vs-JAX log_det_regularization_matrix_term after-measurement following PyAutoArray#631.

• Routing. Under default settings the factory picks the faster measured arm in all six cells. The 512 timing pairs are among the cost model's calibration points (in-sample); the 125 FFT arm was only timed under JAX jit, used as a proxy.
  • 125×125 (extent 83²): direct_conv at 1,405 / 2,947 / 5,526 vertices.
  • 512×512 (extent 341²): FFT at 1,459 vertices, direct_conv at 2,577 and 5,746.
  • The old gate of 60 misrouted the 2,577 cell (nnz/col 106 → FFT, but direct_conv measured 1.29× faster).
  • The cost model's extent-dependent modelled crossover (~309 nnz/col at 83², ~166 at 341²) answers the numba kill-gate question.
• Log-det parity. numba vs JAX differ by ≤ 4.4e-8 at every cell.
  • The 125 original vector went from 2443.22 / 2561.98 to 2468.4463180 on both backends.
• Timings. s/call is carried from the prior laptop measurement (PROGRESS3/PROGRESS4) and is not re-timed.

Campaign: PyAutoPulse#62, task interferometer_large_mask_aris. Library PRs: PyAutoArray#628 / #632 (17caf228), #627 / #631 (0377eeaa), #626 / #630 (7cc71238).

Scripts Changed

• None. This is a results note and a wiki update, with no script changes:
  • results/notes/interferometer_route_gate_aris_2026_10.md (new): the six-cell table, the 125 original-vector pair before and after, settings, SHA and links.
  • wiki/campaigns/interferometer_likelihood.md: the verdict line (fixed gate superseded by the cost model), the ledger and library-PR lines, a new phase row, the gate caveat and a 2026-10-15 journal entry.
• The three scripts that set interferometer_numba_nnz_per_source_max explicitly are unchanged.

Upstream PR

PyAutoLabs/PyAutoArray#632 (MERGED 17caf228; also #631, #630 merged)

Test Plan

• ruff check ., ruff format --check . (no .py added)
• check_wiki.py --check OK (28 campaign files, 52 ledgers linked); check_results_layout.py --check OK; build_readme.py --check, build_dashboard.py --check, build_catalogue.py --check, list_timing_assertions.py --check, check_submits.py --check all current
• Measurements: 6 cells, one process each, default settings, laptop, 9 GB RSS watchdog never fired (peak 1.65–3.04 GB)

Validation checklist (--auto run — plan was not pre-approved)

• Effective level: supervised (header: supervised, cap: feature small → safe); decide-and-flag at ship
• Plan: on the issue (perf: re-measure Aris interferometer routing table under the cost-model gate #419), written at start, unmodified since
• Gate: tests n/a (autolens_profiling has no pytest suite; its PR gate is lint.yml, all of whose checks pass locally, listed above) · smoke n/a (lint.yml's smoke imports scripts; this diff touches no script, only markdown under results/notes and wiki) · review CLEAN (independent Opus reviewer subagent: first pass FINDINGS F1 medium — 512 timing pairs are the cost model's calibration points, so 6/6 is partly in-sample; F2 low-medium — 'no fixed constant serves both masks' was unverified; both resolved in 3237286 by scoping the wording; re-review CLEAN with per-claim dispositions; F3 low, out of scope — the decision-matrix note may still describe gate 60 as current, follow-up) · Heart RED, development-only override (below)
• Human: plan sound in hindsight?
• Human: diff matches plan (no scope creep)?
• Human: merge, amend, or reject — then log the outcome

Decision taken

• Decision: s/call in the table is carried from the prior laptop measurement (Aris PROGRESS3/PROGRESS4, PyAutoArray ccddfba6), clearly marked "prior measurement, laptop, not re-timed". Only routing, predicted t_fft/t_dc and the log-det pair were measured fresh at 17caf228. This was the human's compute decision ("Laptop, routing + log-det only").
• Rejected alternative: re-time all six cells on the laptop. It was shared (load 2–14 today), and the bake-off documents ~2.5× clock drift, so fresh timings would not be cleaner than the paired ones already recorded.
• Revert: delete results/notes/interferometer_route_gate_aris_2026_10.md and revert the wiki hunk.

Heart RED development override

Heart was RED at launch and at ship (2026-10-15): Heart: RED — Release RED; monitoring RED · 10/100 · 415 unresolved; [red] release validation FAILED (stage integrate); [red] PyAutoHands/workflows/Python Version Matrix: failure; [red] PyAutoHeart/workflows/Release Integrate: failure; [red] PyAutoMind/workflows/Lifecycle Drift: failure; [red] PyAutoArray/Tests: red; [red] autofit_workspace_test/Smoke Tests: red; [red] autogalaxy_workspace_test/Smoke Tests: red; [red] autolens_workspace_test/Smoke Tests: red; [red] Release validation: NOT release_ready — v2026.10.15.1.dev81701; [red] Worktree drift: 5 orphan / 0 missing / 4 dirty; [red] autolens_inference: red; [red] autolens_profiling: red; [red] euclid_strong_lens_modeling_pipeline: red (full list on #419).
The live human authorized the development-only override in session 3e3dea38 (AskUserQuestion answer "Yes, override for both (Recommended)"). Permitted: commit, push, PR-open behind the autonomous-ship gate. Merge stays a separate human act (/prm). Heart remains RED for release purposes.

Closes #419

🤖 Generated with Claude Code