test(agent): scripted evals follow only next_steps; the surface over the wire (P5)

Agent reason: the envelope is only worth having if an agent can do its work
by following it. crates/ledger-cli/tests/agent_evals/ holds six scripted
transcripts, no model: onboarding to a first post, a receipt-backed post, a
draft that waits for its receipt and posts when it arrives, a reversal, a
reconciliation, a month-end close. agent_evals_follow_only_next_steps plays
each against a real books server: after books.start a step may call only a
tool (or fetch only a prompt) the previous result's next_steps named, with
the arguments the step gave and the ones the agent decides; setup calls (the
chart and the books an agent finds) are not the agent's moves.

agent_surface.rs holds the over-the-wire tests of the rest: the journal in
words (order, subject and type filters, paging, a bad subject refused),
source corrections through the three tools, the prompts rendered against
their files and refused when malformed, the resource templates advertised
and read, books.start v2 whole and under 8 KiB with the self-report in the
access log only, a new book told to onboard, and a mutating call answering
with the head the journal has and a refusal naming its remedy.

Docs: testing.md's P5 section (every test, and the For-P12 paragraph naming
what a mutant would flip), parity.md (no feature moved; one row), and
architecture.md (the envelope and the guards); mcp-interface.md's overview
sections now describe the surface as built.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HS8WZziDnyxunC23MUxrLp

docs/parity.md

@@ -223,3 +223,18 @@ backdated entries included: defect 6), `zero_net_accounts`,
 `drafts_by_state` and the reconciliation split; reports gained
 `as_of_status` and `strict`; the CPA package is refused while a draft
 dated in it is open (4.8).
+
+### P5: the agent surface (2026-10-29)
+
+No semantic feature moved, so no status changed and the floor stands. P5
+changes how an agent is told what to do, not how a figure is computed: the
+result envelope on every tool result (`next_steps`, `guards`, `warnings`,
+`head`), `books.start` v2, the prompts, the resource templates,
+`books.journal_history` (the journal in words) and source corrections
+(`source_revisions`: `convention.source_revised`, `decision.source_revised`,
+`open_item.source_revised`). One new row, `partial` until a hand-verified
+golden covers it:
+
+| Area | Counterpart (a cross-check, not the spec) | Status | Notes |
+|---|---|---|---|
+| Audit trail in words (`books.journal_history`, the `describe` sentences) | QBO "Audit log", Xero "History and notes" | partial | tests `journal_history_traces_what_the_books_did`, `every_event_type_has_a_sentence`; the viewer keeps its own sentences (the shared describer is the server's) |