PyAutoLabs · Jammy2211 · Aug 21, 2026 · Aug 21, 2026 · Aug 21, 2026 · Aug 21, 2026
diff --git a/dashboard.html b/dashboard.html
@@ -38,7 +38,7 @@
 <body>
 <h1>📋 PyAutoMind Dashboard</h1>
 <p>Every task the Mind is holding. Tap a task's 📋 and its <code>/start_dev</code> command is on your clipboard — paste it into a Claude Code chat to route Claude straight to that task.</p>
-<p class="muted">In flight 4 · Parked 3 · Planned 6 · Backlog 145 · <a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/dashboard.md">markdown version</a></p>
+<p class="muted">In flight 4 · Parked 3 · Planned 6 · Backlog 146 · <a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/dashboard.md">markdown version</a></p>
 <h2>Start here</h2>
 <h3>Highest priority <span class="facets">(filed as high) — showing 12 of 33</span></h3>
 <div class="task"><button class="copy" data-cmd="/start_dev draft/triage/jax_zero_contour.md" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/draft/triage/jax_zero_contour.md">TRIAGE: needs manual review before routing</a> — <span class="facets">medium · safe · high</span></p></div>
@@ -84,9 +84,9 @@ <h2>Planned <a class="mdsrc" href="https://github.com/PyAutoLabs/PyAutoMind/blob
 <div class="task"><button class="copy" data-cmd="/route start the planned PyAutoMind task latent-nan-guard-honest-run — its record is in planned.md" aria-label="Copy the Claude command">📋</button><p><b>latent-nan-guard-honest-run</b></p></div>
 </details>
 <h2>Backlog <a class="mdsrc" href="https://github.com/PyAutoLabs/PyAutoMind/tree/main/draft">markdown version</a></h2>
-<p class="muted">145 filed prompts, not started — sorted most-pickable first (priority, then size).</p>
+<p class="muted">146 filed prompts, not started — sorted most-pickable first (priority, then size).</p>
 <details>
-<summary>bug — 36</summary>
+<summary>bug — 37</summary>
 <div class="task"><button class="copy" data-cmd="/start_dev draft/bug/autoarray/numba_first_call_garbage_psf_weighted_data.md" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/draft/bug/autoarray/numba_first_call_garbage_psf_weighted_data.md">Numba sparse-operator likelihood: first-call garbage / intermittent worker corruption</a> — <span class="facets">autoarray · medium · supervised · high</span></p></div>
 <div class="task"><button class="copy" data-cmd="/start_dev draft/bug/autofit/ep_scale_collapse_basin_cure_or_caveat.md" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/draft/bug/autofit/ep_scale_collapse_basin_cure_or_caveat.md">EP hierarchical parent-scale collapse: cure the basin, or document the</a> — <span class="facets">autofit · too-large · human-required · high</span></p></div>
 <div class="task"><button class="copy" data-cmd="/start_dev draft/bug/autolens/pixelization_eager_vs_jit_divergence.md" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/draft/bug/autolens/pixelization_eager_vs_jit_divergence.md">Investigate eager <code>FitImaging.figure_of_merit</code> vs JIT/step-by-step divergence in rectangular pixelization</a> — <span class="facets">autolens · too-large · supervised · high</span></p></div>
@@ -100,6 +100,7 @@ <h2>Backlog <a class="mdsrc" href="https://github.com/PyAutoLabs/PyAutoMind/tree
 <div class="task"><button class="copy" data-cmd="/start_dev draft/bug/autofit/jax_011_message_log_partition_tuple_shape.md" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/draft/bug/autofit/jax_011_message_log_partition_tuple_shape.md">jax 0.11 breaks beta/gamma message log_partition under jit ('tuple' object</a> — <span class="facets">autofit · small · supervised · medium</span></p></div>
 <div class="task"><button class="copy" data-cmd="/start_dev draft/bug/pyautoheart/script_timing_baselines_orphaned_and_window_filled.md" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/draft/bug/pyautoheart/script_timing_baselines_orphaned_and_window_filled.md">Heart script_timing baselines are orphaned by path moves and filled</a> — <span class="facets">pyautoheart · small · supervised · medium</span></p></div>
 <div class="task"><button class="copy" data-cmd="/start_dev draft/bug/autolens_workspace_test/jax_grad_local_assertions_fail_but_pass_in_ci.md" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/draft/bug/autolens_workspace_test/jax_grad_local_assertions_fail_but_pass_in_ci.md">jax_grad scripts fail assertions locally that PASS in CI</a> — <span class="facets">autolens_workspace_test · medium · supervised · medium</span></p></div>
+<div class="task"><button class="copy" data-cmd="/start_dev draft/bug/autoarray/numba_kernel_shift_axes_swapped.md" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/draft/bug/autoarray/numba_kernel_shift_axes_swapped.md">Numba PSF gathers derive the y/x kernel shifts from the</a> — <span class="facets">autoarray · low · supervised · medium</span></p></div>
 <div class="task"><button class="copy" data-cmd="/start_dev draft/bug/autoarray/pynufft_scipy_pinv2_dev_extra.md" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/draft/bug/autoarray/pynufft_scipy_pinv2_dev_extra.md">PyNUFFT dev extra is incompatible with current SciPy on Python</a> — <span class="facets">autoarray · small · supervised · normal</span></p></div>
 <div class="task"><button class="copy" data-cmd="/start_dev draft/bug/autofit/loggaussian_prior_declares_own_support.md" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/draft/bug/autofit/loggaussian_prior_declares_own_support.md"><code>LogGaussianPrior</code> misreports its own support as <code>(-inf, inf)</code></a> — <span class="facets">autofit · small · supervised · normal</span></p></div>
 <div class="task"><button class="copy" data-cmd="/start_dev draft/bug/autofit/plot_functions_discard_kwargs.md" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoMind/blob/main/draft/bug/autofit/plot_functions_discard_kwargs.md"><code>autofit.plot</code> functions accept <code>**kwargs</code> and silently discard them</a> — <span class="facets">autofit · small · supervised · normal</span></p></div>

diff --git a/dashboard.md b/dashboard.md
@@ -11,7 +11,7 @@ Every task the Mind is holding, on one page: what is in flight, what is parked,
 | [In flight](#in-flight) (`active/`) | 4 |
 | [Parked](#parked) (`parked.md`) | 3 |
 | [Planned](#planned) (`planned.md`) | 6 |
-| [Backlog](#backlog) (`draft/`) | 145 |
+| [Backlog](#backlog) (`draft/`) | 146 |
 
 ## Start here
 
@@ -273,10 +273,10 @@ Scoped but not started; some are not yet prompt files. Full detail in [`planned.
 
 ## Backlog
 
-**145** filed prompts, not started. Each section is sorted most-pickable first (priority, then size).
+**146** filed prompts, not started. Each section is sorted most-pickable first (priority, then size).
 
 <details>
-<summary><b>bug</b> — 36</summary>
+<summary><b>bug</b> — 37</summary>
 
 <details><summary>📋 <a href="draft/bug/autoarray/numba_first_call_garbage_psf_weighted_data.md">Numba sparse-operator likelihood: first-call garbage / intermittent worker corruption</a> — autoarray · medium · supervised · high</summary>
 
@@ -382,6 +382,14 @@ Scoped but not started; some are not yet prompt files. Full detail in [`planned.
 
 </details>
 
+<details><summary>📋 <a href="draft/bug/autoarray/numba_kernel_shift_axes_swapped.md">Numba PSF gathers derive the y/x kernel shifts from the</a> — autoarray · low · supervised · medium</summary>
+
+```
+/start_dev draft/bug/autoarray/numba_kernel_shift_axes_swapped.md
+```
+
+</details>
+
 <details><summary>📋 <a href="draft/bug/autoarray/pynufft_scipy_pinv2_dev_extra.md">PyNUFFT dev extra is incompatible with current SciPy on Python</a> — autoarray · small · supervised · normal</summary>
 
 ```

diff --git a/draft/bug/autoarray/numba_first_call_garbage_psf_weighted_data.md b/draft/bug/autoarray/numba_first_call_garbage_psf_weighted_data.md
@@ -55,3 +55,89 @@ as the regression probe.
   (`_materialize_all`). Worth testing: `cache=False`, numba version pin, and
   whether symptom 2 reproduces with symptom 1 fixed (they may share a cause).
 - Keep the profiling-harness corruption counters as the acceptance test.
+
+## Root cause — confirmed 2026-08-21
+
+Not a numba codegen or caching bug. `psf_weighted_data_from` gathers the
+weight map at `[ip0_y + k0_y + kernel_shift_y, ip0_x + k0_x + kernel_shift_x]`
+with **no bounds check**. numba `@jit()` does not bounds-check array reads, so
+for any unmasked pixel within `kernel_shape // 2` of the array edge the gather
+reads uninitialized heap memory instead of raising `IndexError`. Negative
+indices are unsafe in the same way — they wrap to the opposite edge.
+
+Proof: compiling the shipped source unchanged under `boundscheck=True` raises
+`IndexError: index is out of bounds` for a mask reaching the array edge, and is
+clean for an interior-only mask. With the guard added, the numba output matches
+the zero-padded numpy twin exactly across array sizes and kernel sizes.
+
+This explains **both** symptoms, and explains why the inputs were verified
+identical between call 1 and call 2 — they were; the function reads memory
+*outside* its inputs:
+
+- **Symptom 1** — a cold-cache first call runs right after numba's compilation
+  has churned the heap, so the memory next to the freshly allocated weight map
+  holds compiler garbage (~1e299). Warm cache: no compile, benign neighbour.
+- **Symptom 2** — each forked worker has a different heap layout, so whether
+  the neighbouring memory is poisonous varies per worker and per run. Hence
+  2/8 corrupted in one map and 0/24 in the next.
+
+The `np.isnan` guard was doing real work (masked border is `0/0 = NaN`) but
+never protected the array bounds. The sibling `psf_precision_value_from` was
+already hardened against exactly this — `psf_weighted_data_from` was missed.
+
+The inf suspect (finite image / zero noise passing the `isnan` guard) is **not
+reachable** via the caller: `.native` zeroes both data and noise outside the
+mask, giving `0/0 = NaN`, never `inf`. An inf would require a zero noise value
+*inside* the mask, which is a data-validation error and should stay loud. The
+`isnan` guard is therefore left as-is.
+
+Fix: @PyAutoArray PR #456 (branch `claude/autoarray-numba-psf-garbage-hfxnjv`) — bounds
+guard mirroring the sibling, plus a numba-vs-numpy equivalence regression test
+on an edge-touching mask (fails without the fix, passes with it). Full
+`test_autoarray` suite: 1034 passed, 3 pre-existing pynufft failures unrelated
+to this change.
+
+## Reproduced on the euclid dataset — 2026-08-21
+
+Symptom 1 reproduced exactly, on the real profiling dataset
+(`autolens_profiling`, `dataset/imaging/euclid`, mask radius 3.5", PSF 21x21),
+by calling `psf_weighted_data_from` directly on the masked dataset:
+
+| | `max abs(psf_weighted_data)` | sum |
+|---|---|---|
+| pre-fix (`1c33850`) | **4.901e300** | 1.333e301 |
+| post-fix | **298.312** | 559433.417 |
+
+The bug report's own numbers were "max abs = 4.8e299 on fit #1, 2.98e02 on fit
+#2". The post-fix value **298.31 = 2.98e02** matches the report's *correct*
+value exactly, and the pre-fix value reproduces the uninitialized-memory scale.
+1244 of the 3841 unmasked pixels (32%) drive the gather off the array.
+
+Note on the mask padding — it does **not** protect this path. `apply_mask`
+emits no padding warning and leaves `data.native` and `data.mask` at (71, 71);
+only `derive_mask.blurring_from(allow_padding=True)` pads, to (89, 89), and
+that padded blurring mask is used by the dense convolver, not by the sparse
+numba path. `psf_weighted_data_from` reads the unpadded (71, 71) array via
+`data.mask.derive_indexes.native_for_slim`, so the mask sits flush against the
+array edge and the 21x21 kernel reads past it.
+
+## The acceptance probe is a weak detector — use the direct check instead
+
+`parallel_scaling/pixelization_numba.py` was run at P=2, 24 evals, 2 map
+repeats, cold `NUMBA_CACHE_DIR`, both pre-fix and post-fix. **Both runs
+reported `corrupt_evals_first_map = 0` and `corrupt_evals_steady_maps = [0, 0]`,
+and both had a finite warm-up likelihood.** That is not evidence of no bug: the
+values an out-of-bounds read returns are whatever the allocator left next to
+the weight map, so in that process they happened to be benign. The pre-fix
+warm-up likelihood still drifted from the post-fix one in the 6th decimal
+(5860.175003698866 vs 5860.175922117387) — the same reads, landing on small
+values instead of huge ones. This is exactly the reporter's own intermittency
+(2/8 corrupted in one map, 0/24 in the next).
+
+So the counters can sit at zero on a run where the bug is fully present. Prefer
+the direct `max abs(psf_weighted_data)` check above as the regression probe —
+it is deterministic within a process and reproduces the reported magnitudes.
+
+Split out while fixing this: `draft/bug/autoarray/numba_kernel_shift_axes_swapped.md`
+— both numba gathers derive the y/x kernel shifts from the transposed kernel
+axes, harmless for square kernels but wrong for non-square odd PSFs.
diff --git a/draft/bug/autoarray/numba_kernel_shift_axes_swapped.md b/draft/bug/autoarray/numba_kernel_shift_axes_swapped.md
@@ -0,0 +1,67 @@
+# Numba PSF gathers derive the y/x kernel shifts from the wrong kernel axes
+
+Type: bug
+Target: autoarray
+Repos:
+- @PyAutoArray
+Difficulty: low
+Autonomy: supervised
+Priority: medium
+Status: formalised
+
+Found 2026-08-21 while fixing
+`draft/bug/autoarray/numba_first_call_garbage_psf_weighted_data.md` (the
+out-of-bounds gather in `psf_weighted_data_from`). Split out under the
+one-prompt-one-task rule: separate defect, separate blast radius.
+
+## Symptom
+
+Both numba PSF gathers in
+`autoarray/inversion/inversion/imaging_numba/inversion_imaging_numba_util.py`
+compute their kernel half-widths from the **transposed** kernel axes:
+
+```python
+kernel_shift_y = -(kernel_native.shape[1] // 2)   # shape[1] is x
+kernel_shift_x = -(kernel_native.shape[0] // 2)   # shape[0] is y
+```
+
+at `psf_weighted_data_from` (line ~48) and `psf_precision_value_from`
+(line ~294). The y shift must come from `shape[0]` and the x shift from
+`shape[1]`.
+
+The zero-padded numpy twin
+(`imaging/inversion_imaging_util.py:psf_weighted_data_from`) gets it right and
+is the reference:
+
+```python
+Ky, Kx = kernel_native.shape
+ph, pw = Ky // 2, Kx // 2
+```
+
+## Reachability
+
+Harmless for square kernels (`shape[0] == shape[1]`), which is the common
+case and why no test catches it. It is **not** unreachable: kernels are
+validated as *odd* in each axis, not square — `exc.KernelException("Convolver
+Convolver must be odd")` in `operators/convolver.py:268` and
+`structures/grids/uniform_2d.py:1153` check parity only. A non-square odd PSF
+(e.g. 3x5) therefore mis-centres the gather, sampling the weight map / noise
+map off-centre along both axes.
+
+With the bounds guard now in place the mis-centred reads are clipped rather
+than reading uninitialized memory, so this is a silent wrong-answer bug, not
+a crash or a garbage-value bug.
+
+## Fix
+
+Swap the two right-hand sides in both functions. Fix them **together** — they
+must agree on kernel orientation, and correcting only one would make the
+`psf_weighted_data` and `psf_precision_operator` paths disagree.
+
+## Acceptance
+
+Extend the numba-vs-numpy equivalence test added by the OOB fix
+(`test_autoarray/inversion/inversion/imaging/test_inversion_imaging_util.py::
+test__psf_weighted_data_from__unmasked_pixels_on_array_edge`) to a non-square
+odd kernel (e.g. 3x5). It passes today only because that test uses a square
+kernel; with a non-square kernel the two implementations diverge.