The examples README's annotated session quotes lines the transcript no longer contains

[macanderson]

Problem

examples/README.md's "A complete stdio session (annotated)" walks through the transcript step by step and quotes a JSON line for each step. Those quoted lines are hand-copied, nothing checks them, and they have drifted from examples/full-stdio-session.ndjson.

Concretely, as of PR #138:

• The section used to open with "contains exactly these five lines"; the file had seven lines before that PR and has nine after. (PR feat(wire): carry provenance attestations on the frames envelope #138 replaced the count with a pointer to the file, so the sentence is no longer wrong — but the quoted bodies below it still are.)
• The quoted query and frames lines carry no id, while the file's do (q1), so the walkthrough silently fails to demonstrate correlation, which is the one thing H4 asks a provider to get right.
• The quoted frames line omits content_digest, recorded_at and relations, all of which the file's frames carry, and its token_cost values (41, 23) do not match the file's (32, 18) — so the prose claiming they "sum to 64" describes neither.
• The verify/verified exchange in the file has no step in the walkthrough at all.

The rest of the README is fine; this is one section.

Why it matters

The README says these are "the exact JSON shapes a host and provider exchange — useful when implementing a provider in any language, because you can diff your own output against them." A provider author who diffs against the README rather than the file gets an answer that is wrong in a way schema/validate-examples.py cannot see, because the validator reads the .ndjson and the .json, never the fenced blocks in the README.

Reproduce

python3 -c "
import json
for i, l in enumerate(open('examples/full-stdio-session.ndjson'), 1):
    if l.strip(): print(i, json.loads(l)['type'], json.loads(l).get('id'))
"

then compare against the fenced blocks in examples/README.md.

Constraints

• schema/validate-examples.py already extracts and validates fenced jsonc blocks from `SPEC.md` (section 4 of that script), including a `//`-comment stripper and a placeholder substitution table. The same machinery can read `examples/README.md`'s json blocks — but validating them against the schema only proves they are well-formed envelopes, not that they match the file. The stronger check is equality against the corresponding transcript line.
• The walkthrough deliberately elides some fields for readability, so a byte-equality check would force the prose to carry full frames. Decide whether the section should quote the file verbatim (simplest to enforce) or stay abridged with a weaker check.

Done looks like

• The quoted blocks agree with examples/full-stdio-session.ndjson, including the correlation ids, and the walkthrough covers every message the file contains.
• Whatever agreement is chosen is enforced — a check that fails when the file and the README part company, not a promise to remember. Extending schema/validate-examples.py is the obvious home; it already runs in CI as "schema validates the examples".
• Any number stated in the prose ("sum to 64") is either correct or replaced by something that cannot go stale.