tracer: keep the whole block body, and say why capture failed

[JadenFiotto-Kaufman]

Two findings from the agent stress sweep, both in the with-block capture path
(items 14.2 and 22 in the review notes).

A column-0 comment truncated the trace body

Tracer._parse_block ended its slice at any non-blank line indented back to the
header's column. A comment is non-blank and Python ignores its indentation, so
commented-out code at column 0 — where every editor leaves it — cut the body
short. The truncated slice still parsed as a valid ast.With, so the whole-file
fallback never fired and the statements after the comment silently never ran.

Blanks and comments no longer end the slice. The docstring's "It never returns a
wrong node" was exactly wrong for this case and now says which bound is unsafe
and why.

Before / after on the repro
(tasks/trace-body-capture-and-save/skills/repro_7_col0_comment_truncates_body.py):
NameError: name 'o' is not defined → after comment ran / o = 2.0.

WithBlockNotFoundError carried no message

raise WithBlockNotFoundError() gave the bare class name for at least five
distinct causes — a trace typed into stdin, python -c, an exec'd string not
in linecache, a %%time cell magic, an edit that shifted line numbers — and
the user could not tell them apart. _not_found_message now branches on what the
raise site knows:

• no source recovered at all (stdin, an exec'd string not in linecache) —
  say where a trace body has to live;
• a real file with no with at that line — name the file, line and the text of
  that line, and say the source looks stale (nnsight reads each file once) or
  that the tracer was entered without a with;
• otherwise — name file and line and explain that nnsight compiles the block's
  own body, so line numbers have to match (a %%time cell magic, generated
  source out of step with its linecache entry).

docs/errors/index.md now carries the real message text for all three.

Deliberately not fixed: the stale compiled-body cache itself (BLOCKS/SOURCES
never re-validated), per the review. The second branch only reports that the
source looks stale.

Docs

docs/errors/index.md, plus the two lines my change made false —
docs/errors/with-block-not-found.md and docs/errors/symptom-index.md both
asserted the exception carries no message — and the _parse_block bound and
_not_found_message sentences in docs/developing/tracing-pipeline.md.

Not in this PR

An earlier revision of this branch also restored the thread's trace function
around a block (skip_context clobbers it, __exit__ sets it to None, so a
debugger breakpoint set after a trace never fires and coverage reports every line
after the first block as missed). The maintainer has decided against that change,
and it has been removed in full: skip_context installs the bare no-op and
__exit__ keeps its unconditional sys.settrace(None), exactly as on 0.8.

Testing

PYTHONPATH=<worktree>/src throughout, so the worktree is what ran.

• tests/test_tracing.py: 63 passed (56 before; 7 added). The two comment tests
  fail on origin/0.8 and pass here — verified against a scratch checkout of
  origin/0.8's src. The four _not_found_message tests are new-API and can
  only run here.
• Full suite minus tests/vllm, tests/tp, tests/performance: 1008 passed,
  3 skipped, 1 xfailed. origin/0.8 on the same command: 1001 passed, same
  skips — the difference is exactly the 7 added tests.
• Serialization round-trips re-checked specifically (12 tests across
  test_serialization.py, test_editing.py, test_diffusion.py,
  test_encoder.py): nothing here stores anything new on the Tracer, and
  __getstate__/__setstate__ are untouched relative to 0.8.
• vLLM and tensor-parallel tests were not run: no vLLM in this environment and no
  multi-GPU. Neither item touches those paths.
• Both repros named in the review behave as expected.

Note: pytest tests/test_serialization.py on its own dumps core partway through
in this environment on origin/0.8 too — pre-existing, unrelated, and it does
not happen in the full run.

🤖 Generated with Claude Code