feat(telemetry): a prompt is an axis, and the only one complete by construction - #757

…nstruction

Every breakdown the report carries can read low for a reason about the capture
rather than about the work: a skill the host never named, a run journal that was
never written, an identity nobody declared. `by_prompt` groups on `prompt_id`,
which the transcript reader already resolves for every usage line it stores by
walking each file's own `parentUuid` chain back to the turn that caused the work.

Measured on the built binary against a sandboxed copy of one real session:
1073 of 1073 records carry a `prompt_id`, its 972 subagent records included,
across 12 distinct prompts. The axis reconciles to the same total as every other
breakdown with nothing left in a remainder, on the same session where the step
axis cannot name 96% of the tokens.

A row carries `started_at`, the earliest moment measured inside that prompt,
because an id alone is opaque and the moment is what a person greps for in their
own transcript. The row for records that named no prompt carries neither field,
and is placed last rather than ranked among the prompts: a bucket drawn from many
turns has no size comparable to a single turn's. The text rendering prints the ten
largest and says how many it withheld, since this is the first axis whose
cardinality is unbounded; the envelope always carries them all.

Envelope 12 -> 13, every consumer updated in this commit.

Two gaps closed on the way. The skill's `--axis` enumeration and its
question table both omitted `agent`, which had shipped two versions earlier, so a
person asking which subagent spent the tokens was told the question had no axis
while the binary had been answering it. A guard now reads `ARTEFACT_AXES` from the
module itself and asserts the two lists are equal in both directions. And the e2e
that runs every command the skill names held its own copy of the axis list in a
placeholder regex, so a skill could not name a new axis without that regex being
edited in the same breath; it now expands any `<a|b|c>` by its shape.

Bundle 588.4 -> 590.6 KB, budget raised 590 -> 593 with the measurement recorded
beside it.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01VWNxk63AGKkqE8HRqHLjGp

The axis guard added with `by_prompt` matched the question table with a `\n\n`
terminator, which finds nothing in a checkout whose endings are CRLF - so a suite
green everywhere else failed on Windows alone, naming the first axis it looked
for. Measured: the same regex returns true on the POSIX file and false on the
identical file with CRLF.

The reader normalizes now, once, for every string assertion in this file rather
than only for the one that caught it. Its guard writes a CRLF file to disk and
reads it back through that reader, so a normalization removed from the reader
fails - a string normalized inside the assertion would have passed however the
reader behaved.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01VWNxk63AGKkqE8HRqHLjGp