Phase 1: read Claude Code transcripts into the neutral representation

[JumpTechCode]

The first implementation of the capture port, and the first thing to put the
neutral representation under real load. Claude Code is the reference adapter,
not the substrate: nothing outside this package learns a harness name, and the
architecture test that enforces that stays green.

Two corrections to the representation fell out of building it, both measured
rather than argued:

• No transcript record in any sampled version carries a price — 401 files and
  135,001 records, zero occurrences. Even the reference harness cannot supply
  money from a transcript, so CanSeeCost is not claimed.
• A tool result records that a call errored and never the code it returned,
  so ExitCode became a pointer. Filling it with zero would have reported
  success beside an outcome saying the call failed, and zero is itself a
  meaningful code.

The capability declaration is measured. Cost, the workspace patch, file
mutations and their content are not claimed. Model identity is partial because a
transcript records what served a message and never the alias asked for. Tool
results are partial because large ones are shortened in place. Delegation is
partial because a scan of 601 project-root transcripts found no inline
delegation at all — every one keeps it in sidecar files this adapter does not
read, so an assertion about delegated work resolves to an error rather than to a
false negative.

Against the real store: 300 of 300 transcripts parsed, no refusals, 15,022
turns, 5,112 tool calls matched to 5,112 results — the same pairing the earlier
survey measured. Every committed fixture is synthetic; no transcript content is
committed.

Known defects, filed before merge

An adversarial review after this commit landed found fifteen issues, six of
which are severe enough to name here. They are filed rather than fixed in this
branch, so that the review's evidence is on the record before any of it is
touched:

#25	harness-injected records become human turns — 48.3% turn inflation
#26	CanSeeTokens declared whole while cache token fields go unread
#27	OrderComplete stays true on a compacted session, so the flag is dead
#28	Discover with a Dir query silently matches almost nothing
#29	Model.Provider is filled with a value no transcript states
#30	CanSeeWorkspace declared whole while Revision and Origin are unset

The rest — #31 through #39 — cover wall-clock duration, inline delegation,
sentinel selection, version joining, empty transcripts, two hardening items, a
group of small parser gaps, and one structural gap in the conformance kit
itself: it catches an adapter claiming too little and not one claiming too much.

Four of these are the same class of mistake, and it is worth naming: a
capability declared whole while the fields it governs go unfilled. The
declaration is what decides whether a missing value reads as ERROR or as
FAIL, so an overclaim here manufactures exactly the regression the
representation exists to prevent. #39 is why the kit did not catch them.

The distiller reads this output, so #25 and #27 want fixing before anything is
blessed from it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ew64uotJqtvVDN3twMAVEQ

[codecov]

Codecov Report

❌ Patch coverage is 81.51815% with 56 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/adapter/claudecode/capture/parse.go	83.33%	21 Missing and 12 partials ⚠️
internal/adapter/claudecode/capture/capture.go	77.00%	13 Missing and 10 partials ⚠️

📢 Thoughts on this report? Let us know!