Story-2058 Walkthrough

Change: CE-6 of the Context Engine (Epic-2049, RFD pramaan-functions docs/notes/rfd-context-engine.md §5.1 / §3 / §4). It gives Blake (the registrar) a new capability: from the same document he registers, produce the L1 document_card.v1 — what the document is (type, dates, parties, amounts, 3–12 key assertions) — with a byte-verifiable SourceSpan on every evidential field. Provenance is law: no naked facts.

The core move — "resolution == find"

A card field is only valid if its quote is byte-present in the exact document_pages block it claims (RFD §4, the anti-hallucination gate the >20% kill criterion rides on). The model cannot count bytes, so it never invents char offsets: it names a page + block + a verbatim quote. Code does the rest — a byte-exact block_text.find(quote):

found  → compute char_start/char_end, mint the SourceSpan   (RESOLVED)
absent → drop the field                                    (REJECTED)

So "the model fabricated a quote" and "the span fails to resolve" are the same event, measured deterministically, offline, per-PR. str.find is exact — no whitespace normalization, no case-folding ("a court reads the bytes").

What ships

File	Role
`agents/registrar/context_contracts/`	Vendored CE-1 DocumentCard closure (`DocumentCard`, `SourceSpan`, `CitedValue`, `Card*`, `EnrichmentStamp`), byte-faithful to pramaan-functions `src/shared/context_contracts/` (commit `5b20208`). A CodeZip brain can't import that tree, so we vendor + drift-guard, exactly as Atlas vendors `query_matters_server_params.json`. Do NOT redefine the shape.
`agents/registrar/anchor.py`	Pure plumbing: the page/block-tagged evidence renderer, the byte-exact `resolve_quote_span`, the `SpanGateResult` reject-rate metric, the `sp1:` span-id composite, and `measure_token_inflation`.
`agents/registrar/card_models.py`	The flat generation DTO `RegistrarCardOutput` (shallow, Bug-1357-safe) and `assemble_document_card` — the bridge that byte-resolves every quote into the canonical span-carrying card, dropping what won't resolve.
`agents/registrar/prompts/card.py`	Blake's card system prompt (PROMPTING.md: teach-by-contract, recursive injection fence, provenance-is-law, self-check-before-emit) + the anchored user-prompt builder.
`agents/registrar/agent.py`	`_card_agent` + `extract_document_card` — the fail-closed producer that NEVER raises.
`agents/registrar/runtime.py`, `events.py`	A new `card` mode that emits a `document.card` effect the broker persists (CE-7 ingest boundary). Off the classify turn.
`agents/registrar/evals/card_golden_corpus.jsonl`, `evals/card_gate.py`	The kill-criterion harness: a golden Indian-legal corpus + the deterministic reject-rate gate and the periodic live measurement.

Failure isolation — card failure never blocks registration

The card is a separate pass, not part of the interactive classify turn (RFD §5.1; matches CE-7's separate SQS enrich kick). Two guarantees, both test-asserted:

extract_document_card is fail-closed — any producer error (model stall, timeout, missing provider, bad output) returns a typed status="error" result, it never raises.
The classify/registration path never references the card producer; a card-mode run that fails still completes with run.completed (never run.failed), so the enrich kick returns and the registration consumer is never blocked.

Acceptance criteria → proof

AC	Proof
card failure never blocks registration apply (test-asserted)	`test_registrar_card_isolation.py` — extract never raises (provider-down, model explodes); classify never calls the producer; card mode completes on failure, never `run.failed`.
anchored-input token inflation measured	`anchor.measure_token_inflation` + `test_anchored_input_token_inflation_is_measured` — flat vs anchored token counts + ratio, serialized for the record.
< 20% span rejects on golden corpus (kill criterion)	`test_registrar_card_golden_corpus.py` over `card_golden_corpus.jsonl` — well-formed corpus resolves at 0% (far under 20%); fabricated cases ARE rejected (the gate has teeth). Live Blake reject-rate is `evals.card_gate.measure_live_reject_rate` (periodic, needs a key).

Verification

uv run --package registrar pytest tests -q           # 270 passed
uv run ruff check agents/registrar/                  # All checks passed!
python scripts/check_brain_egress.py                 # brain-egress check: clean
python scripts/check_prompt_leakage.py               # prompt-leakage check: clean
uv run pytest tests/ -q -m unit                      # 170 passed (governance)

Brain stays config+brain only (pramaan-docs ADR-110 §6): the card producer holds no DB/DSN, calls no tool, and its only egress is the inference provider. The context_contracts package is pure (no sqlalchemy/boto3). The broker-side card persistence (CE-7) and the registration-consumer enrich kick are downstream stories.