Rename doom_* noise distributions to neutral names

The five `doom_*` input distributions defined in
`scripts/measure_op_noise.py` were named after the doom-graph callsites
that originally motivated them. After Step E, those callsites live in
the sibling `torchwright_doom` repo; the distributions still encode the
input-range characterizations used to calibrate piecewise ops, but the
names and descriptions pointed at deleted code (Option A from inventory).

Renames (chosen to match the file's existing `<op>_<range>_<details>`
style — e.g. `log_4decades_001_100`, `compare_uniform_pm80`):

- `doom_reciprocal_wall`     → `reciprocal_03_200`
- `doom_reciprocal_sorted`   → `reciprocal_01_50_step01`
- `doom_diff_trig`           → `diff_trig_nonuniform`
- `doom_diff_vel`            → `diff_vel_nonuniform`
- `doom_atan_cross_dot`      → `atan_cross_dot_nonuniform`

Description rewrites (drop the `mirrors the X callsite at
torchwright/doom/...:line` tail; keep the input-range characterization).
The `compare_near_thresh_05` and `linear_bin_index_tex_col_16`
descriptions also had similar dead-callsite tails — scrubbed inline.

Regenerated by `make measure-noise`:
- `docs/op_noise_data.json` (commit SHA + new names)
- `docs/numerical_noise.md` (auto-rendered)
- per-op `.. noise-footer::` blocks in `torchwright/ops/{arithmetic,
  logic}_ops.py`

Hand-edited `docs/numerical_noise_findings.md`:
- Renamed all distribution references to the new names.
- Renamed 'Input-amplification hazards (the Phase E trap, now partially
  mitigated)' to 'Input-amplification hazards in gate ops'; dropped the
  'Worked Phase E example (historical)' subsection (postmortem
  pointer was deleted in Step E).
- Dropped the 'Wired into DOOM' subsection of the `log` finding —
  documented how DOOM used log/exp on `torchwright/doom/stages/render.py`.
- Dropped the E1M1/`inv_scale` worked example from the
  `compare(x, 0)` finding; kept the general working-rule.
- Renamed 'Known gap: reciprocal sorted callsite' to 'Known gap:
  reciprocal at step=0.1'; dropped the `thinking_wall.py:1196`
  callsite reference and 'Latent production implication' paragraph.
- Deleted the entire 'Production call-sites' section: every row was
  a deleted `torchwright/doom/stages/...` path.

CLAUDE.md:
- Removed the 'which DOOM call-sites each distribution covers' clause
  from the Numerical-noise section preamble.
- Removed the 'Keep the call-site cross-reference table in sync when
  you add, rename, or remove a `doom_*` distribution' bullet from the
  measure-noise workflow.

Also scrubbed one stale ops-docstring mention: `log_abs`'s docstring
example 'coord · inv_scale in DOOM normalization' → 'signed_coord ·
positive_scale' (matches the parallel rewrite in the findings doc).

Verified: full `make test` passes — 635 passed, 3 skipped.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: physicsrob

CLAUDE.md

@@ -167,12 +167,12 @@ is the canonical source; `docs/numerical_noise.md` and the per-op docstring
 footers are **generated** from that JSON. Never edit them by hand.
 
 Commentary on the measurements — which numbers are by design, which deserve
-investigation, and which DOOM call-sites each distribution covers — lives in
-`docs/numerical_noise_findings.md`. That file is **hand-written and is
-Claude's responsibility to keep current**. The measurement pipeline does not
-regenerate it and the consistency test does not enforce it; synthesizing
-findings from the raw numbers is an interpretive task that belongs to the
-agent running the workflow, not to the script.
+investigation — lives in `docs/numerical_noise_findings.md`. That file is
+**hand-written and is Claude's responsibility to keep current**. The
+measurement pipeline does not regenerate it and the consistency test does
+not enforce it; synthesizing findings from the raw numbers is an
+interpretive task that belongs to the agent running the workflow, not to
+the script.
 
 Specifically, whenever you run `make measure-noise` — whether because you
 added an op, edited one, or widened a distribution — you must:
@@ -183,8 +183,6 @@ added an op, edited one, or widened a distribution — you must:
    entry in `docs/numerical_noise_findings.md` or add a new one.
 3. Remove findings that no longer apply (e.g., if a bound is tightened and
    the previously-flagged number is now within expectations).
-4. Keep the call-site cross-reference table in sync when you add, rename, or
-   remove a `doom_*` distribution.
 
 Regenerate the auto-artefacts with:
 

docs/numerical_noise.md

@@ -1,14 +1,14 @@
 # Numerical noise reference
 
 *Generated by `scripts/measure_op_noise.py`; do not edit by hand.*
-*Measured at commit `3829b69` on `2026-04-26`.*
+*Measured at commit `c234ba4` on `2026-04-29`.*
 
 Every piecewise-linear op in `torchwright/ops/` is measured on one or more
-named input distributions that mirror its production use in
-`torchwright/doom/`. The canonical data lives in `docs/op_noise_data.json`;
-this file is regenerated from that JSON by `make measure-noise`.
+named input distributions characterising its expected input ranges.
+The canonical data lives in `docs/op_noise_data.json`; this file is
+regenerated from that JSON by `make measure-noise`.
 
-For commentary on the numbers — which are expected, which warrant investigation — and for the DOOM call-site cross-reference, see the hand-written `docs/numerical_noise_findings.md`.
+For commentary on the numbers — which are expected, which warrant investigation — see the hand-written `docs/numerical_noise_findings.md`.
 
 ## Summary
 
@@ -28,15 +28,15 @@ For commentary on the numbers — which are expected, which warrant investigatio
 | `linear_bin_index` | `torchwright.ops.arithmetic_ops` | 1 | 0.9854 | `linear_bin_index_tex_col_16` |
 | `log` | `torchwright.ops.arithmetic_ops` | 0.003026 | 0.003721 | `log_6decades_wide` |
 | `log_abs` | `torchwright.ops.arithmetic_ops` | 0.000699 | 0.1428 | `log_abs_3decades_pm100` |
-| `low_rank_2d` | `torchwright.ops.arithmetic_ops` | 0.0651 | 0.3314 | `doom_atan_cross_dot` |
+| `low_rank_2d` | `torchwright.ops.arithmetic_ops` | 0.0651 | 0.3314 | `atan_cross_dot_nonuniform` |
 | `max` | `torchwright.ops.arithmetic_ops` | 3.815e-06 | 1.953e-06 | `minmax_uniform_pm50` |
 | `min` | `torchwright.ops.arithmetic_ops` | 3.815e-06 | 1.953e-06 | `minmax_uniform_pm50` |
 | `mod_const` | `torchwright.ops.arithmetic_ops` | 0 | 0 | `mod_integers_0_100_by7` |
 | `multiply_2d` | `torchwright.ops.arithmetic_ops` | 0.0625 | 2.15 | `multiply_uniform_pm10` |
 | `multiply_integers` | `torchwright.ops.arithmetic_ops` | 0 | 0 | `multiply_integers_0_10` |
 | `piecewise_linear` | `torchwright.ops.arithmetic_ops` | 0.25 | 231.1 | `parabola_0_10_step1` |
-| `piecewise_linear_2d` | `torchwright.ops.arithmetic_ops` | 7.778 | 0.6024 | `doom_diff_trig` |
-| `reciprocal` | `torchwright.ops.arithmetic_ops` | 0.0009137 | 0.1686 | `doom_reciprocal_wall` |
+| `piecewise_linear_2d` | `torchwright.ops.arithmetic_ops` | 7.778 | 0.6024 | `diff_trig_nonuniform` |
+| `reciprocal` | `torchwright.ops.arithmetic_ops` | 0.0009137 | 0.1686 | `reciprocal_03_200` |
 | `signed_multiply` | `torchwright.ops.arithmetic_ops` | 0.06248 | 2.15 | `signed_multiply_pm10` |
 | `square` | `torchwright.ops.arithmetic_ops` | 0.25 | 231.1 | `square_unsigned_0_10` |
 | `square_signed` | `torchwright.ops.arithmetic_ops` | 0.25 | 276.7 | `square_signed_pm10` |
@@ -309,7 +309,7 @@ End-to-end composite (~6 MLP sublayers). Error compounds across `reciprocal`, `s
 
 ### `linear_bin_index_tex_col_16`
 
-16-bin quantisation of x ∈ [0, 64] with x_min=0, x_max=64 — mirrors texture-column bucketing in the RENDER stage.
+16-bin quantisation of x ∈ [0, 64] with x_min=0, x_max=64.
 
 | Metric | Value |
 | --- | --- |
@@ -379,9 +379,9 @@ log_abs(x) for x ∈ [-100, 100] with min_abs=0.1, max_abs=100 (3 decades on |x|
 
 Rank-3 SVD of `atan(cross/dot)` on a non-uniform grid. Worst-cell error is bounded by σ_{K+1}, the first truncated singular value — verify against the SVD computed in `tests/ops/test_low_rank_2d.py:100`.
 
-### `doom_atan_cross_dot`
+### `atan_cross_dot_nonuniform`
 
-atan2(cross/dot) over [-2, 2] × [0.5, 10] on non-uniform grid — mirrors the `atan_front_*` callsite at `torchwright/doom/stages/sorted.py:549` (rank-3 SVD).
+atan2(cross/dot) over [-2, 2] × [0.5, 10] on a non-uniform grid — rank-3 SVD via `low_rank_2d`.
 
 | Metric | Value |
 | --- | --- |
@@ -512,9 +512,9 @@ f(x)=x² on [0, 10] with integer breakpoints — probes generic `piecewise_linea
 
 Triangulated-grid lookup with non-uniform breakpoints. On non-uniform grids `piecewise_linear_2d` uses a constrained least-squares fit that can oscillate in cell interiors — `low_rank_2d` is preferred there.
 
-### `doom_diff_trig`
+### `diff_trig_nonuniform`
 
-wall-geometry product over DIFF_BP × TRIG_BP — mirrors the `sort_ey_cos`/`sort_ex_sin` family of calls at `torchwright/doom/stages/thinking_wall.py:1115` and neighbours.
+Two-input product over DIFF_BP × TRIG_BP — DIFF in [-40, 40] (non-uniform), TRIG in [-1, 1] (non-uniform).
 
 | Metric | Value |
 | --- | --- |
@@ -527,9 +527,9 @@ wall-geometry product over DIFF_BP × TRIG_BP — mirrors the `sort_ey_cos`/`sor
 | p99 rel error | 0.35 |
 | Worst input | `a=22.22`, `b=-1` |
 
-### `doom_diff_vel`
+### `diff_vel_nonuniform`
 
-collision product over DIFF_BP × VEL_BP — mirrors the `p_dx_ey`/`p_dy_ex` family at `torchwright/doom/stages/thinking_wall.py:1014`.
+Two-input product over DIFF_BP × VEL_BP — DIFF in [-40, 40] (non-uniform), VEL in [-0.7, 0.7] (non-uniform).
 
 | Metric | Value |
 | --- | --- |
@@ -544,11 +544,11 @@ collision product over DIFF_BP × VEL_BP — mirrors the `p_dx_ey`/`p_dy_ex` fam
 
 ## `reciprocal`
 
-Geometric breakpoint spacing keeps relative interpolation error roughly constant across the range. Measured against a reciprocal graph whose `(min_value, max_value, step)` matches its production callsite: `(0.3, 200.0, 1.0)` for `doom_reciprocal_wall` (`torchwright/doom/stages/wall.py:431`).
+Geometric breakpoint spacing keeps relative interpolation error roughly constant across the range. Measured against a reciprocal graph with `(min_value, max_value, step) = (0.3, 200.0, 1.0)`.
 
-### `doom_reciprocal_wall`
+### `reciprocal_03_200`
 
-1/x over [0.3, 200.0], step=1.0 — mirrors the `inv_abs_num_t` callsite at `torchwright/doom/stages/wall.py:339`.
+1/x over [0.3, 200.0], step=1.0.
 
 | Metric | Value |
 | --- | --- |