Phase 1: the neutral session representation and the capture port (ADR-0006)

[JumpTechCode]

Second of three Phase 1 records. Adds internal/assay (the harness-neutral
session representation), internal/port (the capture interface over it), and
internal/adapter/conformance (the shared kit plus a second implementation).

The measurement changed the design

Three independent adversarial reviews all pointed at the same gap: the evidence
for several decisions was already on this machine and had never been read. It
was run, and it moved three decisions.

Measured	Consequence
54 of 2,459 local transcripts used more than one model; 3 used three	the model belongs to a turn, not a session
fork-context-ref records carry parentSessionId	session lineage keeps a parent pointer, measured rather than inferred
file-history-snapshot/-delta carry snapshot, backup, trackingPath	file changes carry content, without which patch-similarity is uncomputable

The first survey had gathered the model set and checked the size of the
working-directory set sitting beside it. The other two record types it had
classified as bookkeeping and skipped.

The part that has teeth

Observability is declared per adapter and the declaration is enforced, not
documented. A capability governs every field; a test fails if any field is
governed by none; an adapter that reports something it did not declare fails
the kit. Both were verified by planting violations:

• a package-level var in the leaf → rejected, naming the file
• a new SidechainUID field on Session → rejected, naming the field and what to do

Partial observation is its own state. Compaction and truncation are ordinary,
and a capability declared whole when it is partial fails an assertion on
evidence that was never complete — the manufactured regression the whole
mechanism exists to prevent.

A tool result carries how the call ended, not whether it "failed". A refused
call, a cap hit, and a call that never returned are not regressions, and one
boolean cannot say so. The exit code is kept because test runners distinguish
"the tests failed" from "what you named does not exist".

The second implementation

Shaped deliberately unlike the harness measured first: no turn structure,
consecutive agent turns, a model change mid-stream, delegated work as
unreferenced siblings, tokens without money, pre-split commands, no version
control. A merely poorer stub would have exercised nothing — leaving fields
empty satisfies any shape, and shape is where single-vendor assumptions hide.

Stated rather than claimed

• The turn is a reconstruction. Neither harness examined stores turns; both
  are flat record streams. Where an adapter draws boundaries determines any
  turn-counting metric.
• Workspace assumes a local filesystem and a content-addressed VCS, said out
  loud rather than hidden behind neutral names.
• Cost is close to a single-vendor feature. Relative comparisons are
  token-denominated; absolute caps apply where a price is actually reported.
• Neutrality is not demonstrated here. One measured harness, a paper reading
  of a second, and a reference written by the same hand. A check, not a proof.

Also records — without correcting — that S2's two published subagent-sidecar
counts disagree and the store has grown too much to settle which was right.
Choosing one of two irreproducible numbers would be worse than saying so.

make verify green: coverage 88.1%, 0 lint issues, 0 vulnerabilities.

[codecov]

Codecov Report

❌ Patch coverage is 85.62992% with 73 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/adapter/conformance/verify.go	32.35%	26 Missing and 20 partials ⚠️
internal/adapter/conformance/coverage.go	84.37%	15 Missing and 5 partials ⚠️
internal/assay/enum.go	96.15%	5 Missing ⚠️
internal/adapter/conformance/reference.go	97.59%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!