fix(adapter): close the review findings against the Claude Code capture

[JumpTechCode]

Twelve of the thirteen issues filed against the first capture adapter after it
shipped. Each was reproduced before it was fixed, and the effect of every fix is
measured against a local store of 733 project-root transcripts.

Measured, before and after

	before	after
input tokens across the store	3,888,052	12,470,443,100
transcripts with a negative Wall	2	0
working directories a Dir query finds	66 of 300 sampled	185 of 185, 0 missed
transcripts that fail to load	0	0
sessions withholding the order claim	87	88

The last row is the orphan-result rule finding one file the compaction markers
did not.

Wrong numbers reaching a verdict

• capture(claudecode): CanSeeTokens is declared whole while most input tokens go unread #26 InputTokens read one of three fields. cache_read_input_tokens and
  cache_creation_input_tokens carry 99.97% of the input on the measured store,
  so a token-band assertion ran against a figure three orders of magnitude
  small, and a change in the harness's caching — one of the things an exam is
  re-administered to detect — was invisible in it.
• capture(claudecode): Wall duration saturates to 292 years, or runs negative #31 Wall was last minus first, unguarded. A missing timestamp saturated
  it to time.Duration's maximum; records out of order, which 445 of 728
  transcripts have, made it negative on two. It is now the span between the
  timestamps that exist, and assay.Usage says what that does and does not
  measure.
• capture(claudecode): Model.Provider is filled with a value no transcript states #29 Model.Provider was set to a constant. No record names a provider,
  and the harness writes the same transcript against three backends.
• capture(claudecode): Discover with a Dir query silently matches almost nothing #28 Discover with a Dir query missed most directories: the store
  replaces every character outside [A-Za-z0-9-], this replaced only the
  separator. The comparison was also strings.Contains, so /work matched a
  project named for /workspace/other; it is equality now.

Sentinels and claims

• capture(claudecode): an over-long line reports ErrMalformed where the source is intact #33 An over-long line reported ErrMalformed — damage — for an intact
  source. It is this adapter's own limit, so ErrUnsupported. A null or
  type-less line, which unmarshalled into a zero record and was refused as
  unknown record type "", is now malformed with a reason that names the cause;
  a leading byte order mark is no longer refused.
• capture(claudecode): an empty transcript loads as a complete zero-turn session #35 An empty or bookkeeping-only file loaded as a complete zero-turn
  session — indistinguishable from a session in which nothing happened, and
  claiming to know the complete order of that nothing.
• capture(claudecode): CanSeeWorkspace is declared whole while Revision and Origin are never set #30 CanSeeWorkspace was declared whole while four of its six governed
  fields are never filled. It is partial, with the reason stated.
• capture(claudecode): Fidelity.Version joins several versions into one string #34 Fidelity.Version joined releases into "2.1.219,2.1.220", which the
  equality check the field exists for can never match. It names the release the
  session's last record was written by; Verified still covers all of them.

Structure

• capture(claudecode): inline delegated work loses its results and its tokens #32 Delegated work never had its results paired or its tokens counted: the
  sidechain branch returned before both. Sub-agent usage now lands on the
  delegation rather than being dropped, and capture.go drops the claim that
  the inline form was read correctly.
• capture(claudecode): tool-result pairing is safe by accident, not by construction #36 Pairing keyed on a pointer into a turn's tool slice. Safe today,
  orphaned in silence by any later change that appends a call to a stored turn.
  Keyed by position, resolved after the file is read.
• capture(claudecode): assorted parser gaps found by census #38 An orphan result now withdraws the order claim rather than being
  discarded; an unknown record type is truncated before it is formatted into an
  error. The other four bullets are documented rather than changed:
  attachment records keep their payload on a key this adapter does not read;
  started and result turn out to live only in workflow journals, which are
  not transcripts, so refusing them with ErrUnsupported is correct; fallback
  blocks record a model substitution on a record that becomes no turn, and there
  is nowhere neutral to put it; the pending map is bounded by one file.
• capture(claudecode): each message body is JSON-parsed twice #37 Each message body was unmarshalled twice, on lines that reach 4.2 MiB.
  One walk now produces the turn and the identifiers.

Left open

#39 — the conformance kit catches an adapter reporting what it cannot see,
and cannot catch one declaring a capability it never fills. The comment claiming
otherwise is corrected in coverage.go and verify.go. The check itself needs
a fixture per capability known to carry the evidence for it, which does not
exist yet, so the issue stays open rather than being closed by a comment.

Tests

Fifteen new cases, each written first and watched fail. All fixtures are
synthetic; no transcript is committed. make verify passes: race tests,
coverage 88.9%, lint 0 issues, govulncheck clean.

Closes #26, #28, #29, #30, #31, #32, #33, #34, #35, #36, #37, #38

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ew64uotJqtvVDN3twMAVEQ

[codecov]

Codecov Report

❌ Patch coverage is 98.70130% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/adapter/claudecode/capture/parse.go	98.57%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!