Sweep remaining broken-pointer doom citations

Five spots cited specific deleted artifacts (deleted test files, deleted
flows, deleted callsite tables). Each carried a citation that no reader
could resolve.

- `torchwright/compiler/components/attn.py:22` — dropped the
  `tests/doom/test_thinking_readback.py::test_emit_continuous_roundtrip`
  citation. The surrounding paragraph already explains the failure mode
  it demonstrated.
- `torchwright/compiler/export.py:953` — dropped 'dev-mode DOOM play'
  from 'Used for dev-mode DOOM play and compiler-output tests…'. The
  dev-mode play flow lived in deleted code; compiler-output tests are
  the only surviving caller.
- `torchwright/debug/noise.py:219` — dropped 'DOOM callsites' from the
  noise-footer generator's docstring; after the bucket-3 distribution
  rename, distributions no longer carry callsite descriptions, so the
  warning was misleading. Kept the rule (per-distribution context lives
  in the markdown reference, not in op docstrings).
- `tests/ops/test_piecewise_linear_2d.py:4` — dropped 'it was only
  exercised indirectly through the DOOM renderer tests' from the
  module docstring. Those tests don't exist here. The remaining
  description (what the tests verify) stands on its own.
- `CLAUDE.md` § 'probe_compiled' — rewrote the 'Interpreting atol'
  paragraph. Was: 'For the DOOM renderer (deep op chains, values in
  the 10^4 range), atol=500 is the empirical floor — see the atol
  comment in tests/debug/test_probe.py:test_probe_clean_on_v2_box_room'.
  That test was deleted in the Step E aftermath. Replaced with a
  general rule (op-chain depth and value-range magnitude both push
  the floor up; tighten on precision wins).

Three flavor mentions ('e.g. the DOOM renderer', 'e.g. DOOM's
rows_per_patch', 'typical DOOM pattern') in the same files were left
alone — they cite DOOM as a worked example without broken pointers,
which is fine per the user's standing policy.

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

Author: physicsrob

CLAUDE.md

@@ -355,11 +355,14 @@ one call:
                          d=2048, d_head=32, atol=500.0)
 
 **Interpreting `atol`.**  The tolerance must account for accumulated
-piecewise-linear approximation error through the graph.  For a
-shallow graph (few ops, small value range), `atol=1e-3` is
-appropriate.  For the DOOM renderer (deep op chains, values in the
-10^4 range), `atol=500` is the empirical floor — see the atol
-comment in `tests/debug/test_probe.py:test_probe_clean_on_v2_box_room`.
+piecewise-linear approximation error through the graph.  Both op-chain
+depth and value-range magnitude push the floor up: a shallow graph
+with values near 1 can use `atol=1e-3`; a deep chain operating on
+values in the `10^4` range may need `atol` of several hundred to
+absorb fp32 accumulation in the matmul-based residual writes.  Set it
+just above the largest legitimate divergence the oracle comparison
+should tolerate, and tighten it whenever you land a precision
+improvement.
 
 ## probe_residual — layer-by-layer node values
 

tests/ops/test_piecewise_linear_2d.py

@@ -1,10 +1,8 @@
 """Unit tests for piecewise_linear_2d.
 
-The 2D primitive had no dedicated tests — it was only exercised
-indirectly through the DOOM renderer tests.  These tests verify
-correctness in isolation: exact at grid points, bounded interpolation
-error, boundary clamping, non-uniform grids, and compiled-vs-oracle
-agreement via probe_graph.
+These tests verify correctness in isolation: exact at grid points,
+bounded interpolation error, boundary clamping, non-uniform grids, and
+compiled-vs-oracle agreement via probe_graph.
 """
 
 import math

torchwright/compiler/components/attn.py

@@ -18,8 +18,7 @@
 # Why it matters here: cancel heads (V=identity, O=-identity) rely on
 # attn_out + skip = 0 algebraically.  A 1-LSB perturbation at q≈22855
 # leaves a ~1/512 leak in the residual column, which propagates through
-# the encoder and flips Gray-code bits at boundary q values — see
-# tests/doom/test_thinking_readback.py::test_emit_continuous_roundtrip.
+# the encoder and flips Gray-code bits at boundary q values.
 # Same shape applies to compute_linear (Linear emulated as self-attn).
 _SDPA_BACKEND = [SDPBackend.MATH]
 

torchwright/compiler/export.py

@@ -950,10 +950,9 @@ def add(op, ins, outs, **attrs):
 # In-process headless callable: a thin adapter over HeadlessTransformer.compute()
 # that presents the same (inputs) -> outputs interface as OnnxHeadlessModule.
 #
-# Used for dev-mode DOOM play and compiler-output tests that don't need an
-# ONNX round-trip.  For production inference, export with
-# compile_headless_to_onnx instead — that path streams weights and runs
-# under onnxruntime.
+# Used for compiler-output tests that don't need an ONNX round-trip.  For
+# production inference, export with compile_headless_to_onnx instead —
+# that path streams weights and runs under onnxruntime.
 # ---------------------------------------------------------------------------
 
 

torchwright/debug/noise.py

@@ -216,8 +216,8 @@ def render_footer_block(
 
     The footer is deliberately distribution-agnostic: it reports the worst
     observed abs and rel error across every distribution the op was measured
-    on. Project-specific context (DOOM callsites, distribution names) lives
-    in ``docs/numerical_noise.md``, not in the op's docstring.
+    on. Per-distribution context lives in ``docs/numerical_noise.md``, not
+    in the op's docstring.
     """
     body = indent + "   "
     rel_txt = "n/a" if _is_nan(max_rel_error) else _fmt(max_rel_error)