Decide context/resolve before the freeze: specify it or cut reference representations — and define B3 for compact/reference frames

[macanderson]

Capabilities.resolve is advertised (schema/contextgraph-envelope.schema.json:373-376, contextgraph-types/src/capability.rs:133), and compact/reference representations require it (representations_consistent, capability.rs:155) — but there is no resolve/resolved envelope in the schema oneOf, no wire shape, no host API (contextgraph-host/src/compose.rs:85 — "a later phase"), and no conformance check. ADR 0005 (line 77) confirms context/resolve is deferred.

This is exactly the dead-capability-surface anti-pattern ADR 0004 purged (upsert/subscribe/filters): a frozen 1.0 would ship reference frames that no specified mechanism can rehydrate.

The budget half. SPEC.md:250 mandates token_cost = ceil(utf8_byte_length(content)/4); a reference frame has no content, and the code silently decided absent ⇒ 0 (contextgraph-types/src/frame.rs:397-399). B1 budget honesty therefore assigns zero cost to frames whose resolved content is unbounded — a budget hole the moment resolve exists. (Also schema:580's tokenizer_ref comment contradicts ADR 0003's Option A.)

Options (mirroring ADR 0004's discipline)

A. Specify context/resolve pre-freeze: envelope pair, correlation, error codes, and B3 semantics for every representation.

B. Cut reference representation + Capabilities.resolve from 1.0; keep a sketch in docs/sketches/ for a 1.x additive minor. compact may survive if its B3 rule is written down.

Either way, B3 must be restated for every representation the frozen spec ships.

[macanderson]

Data point for this decision from #44 (ADR 0006): the prompt-ingest provider is now a concrete, in-tree consumer of compact/reference representations, and it required taking a position on resolve before this issue is settled.

What it does today, and why it's honest without the context/resolve wire method:

• It serves compact frames by default and advertises representations: [full, compact, reference] + resolve: true (forced by representations_consistent).
• The working rehydration path is a re-query with representation_preferences = [full]: IngestProvider::query returns the full-content frame straight from its content-addressed store, with honest token_cost/content_digest recomputed for the full inline bytes (§B3). So resolve: true is backed by a real, tested callable path — unlike the ADR 0004 dead flags — even though the dedicated wire method doesn't exist yet.
• B3 for compact/reference (this issue's second half) is handled concretely: token_cost is recomputed per representation over the exact inline bytes emitted (compact = distilled rendering, reference = 0), and canonical_token_cost carries the full-source cost. contextgraph-conformance/tests/ingest_conformance.rs asserts this holds across all three representations.

So whichever way this goes — specify context/resolve or cut reference — the ingest provider is a ready reference implementation of the consumer side. If context/resolve is specified, IngestProvider overrides the trait default trivially (the store already answers by content_ref.uri / canonical hash). If reference is cut, the provider drops to [full, compact] with no other change.

[macanderson]

Folded from the #27 adaptive-context reconciliation (#27 delta table, rows D3/C3)

The build prompt specifies context/resolve: an opaque content_ref in → resolved content out, with the host verifying the canonical_content_hash, and a typed failure taxonomy (reference_not_found, reference_expired, scope_denied, sharing_denied, consent_required, content_hash_mismatch, unsupported_representation). That's the resolve design this issue asks to "specify or cut."

Pair it with the B3 half this issue already raises: a reference frame has no inline content, so token_cost = ceil(bytes/4) needs a written rule (the code's silent absent ⇒ 0 is a budget hole once resolve exists). The build prompt carries token_cost/canonical_token_cost/tokenizer_ref for exactly this.

[macanderson]

Re-verified 2026-07-26 against origin/main @ 1ee9b1c. Closing — PR #60 decided it.

context/resolve: decided as a scoped Option B. SPEC §6.4.1 now states that context/resolve is not defined in contextgraph/1.0, and capabilities.resolve is demoted to an explicit forward-declaration rather than a negotiated-but-unimplemented method — which is exactly the dead-surface anti-pattern ADR 0004 purged elsewhere, now closed off here too.

B3 for compact/reference frames: answered per representation at SPEC P4, and enforced by budget-honesty rather than left to convention.

One residual contradiction to carry into #49 (or a one-line schema chore): the schema's tokenizer_ref $comment still contradicts ADR 0003's exact-equality rule. It is a comment, not a constraint, but it is the kind of stale normative-sounding text the freeze should not inherit.