Skip to content

core: close the evaluator union and package session bridge - #454

Closed
lemone112 wants to merge 1 commit into
agent/weak-session-ownerfrom
agent/mixed-evaluator-package-bridge
Closed

core: close the evaluator union and package session bridge#454
lemone112 wants to merge 1 commit into
agent/weak-session-ownerfrom
agent/mixed-evaluator-package-bridge

Conversation

@lemone112

@lemone112 lemone112 commented Jul 22, 2026

Copy link
Copy Markdown
Collaborator

Summary

  • Replace the single-evaluator production seam with one code-owned closed evaluator union: exact encoded-sRGB8 identity and WCAG 2.2. Dispatch is a direct match with concrete evidence variants; no trait object, callback, client registry, or threshold dialect reaches the evaluation loop.
  • Add one concrete package projection over the generation-bound Core Session: opaque owner/session, closed lifecycle and error unions, canonical output slots, borrowed certificates, and total Set | Remove | Hold operations.
  • Add a schema-ordered observation path for the packed package boundary. It uses one reusable scratch vector and one canonical sort, constructs no keyed port-binding vectors, and performs allocation-free exact replay after scratch growth.
  • Acquire the weak owner lease before either observation admission path and map OwnerExpired explicitly.
  • Preserve the single raw observation owner and shared transactional commit path for keyed and schema-ordered admission.
  • Rebind the fail-closed point-support proof to the exact changed semantic cone and ratchet the canonical WASM budget to the measured 376832-byte artifact from Actions run 29966828219, with zero arbitrary headroom. Gzip remains diagnostic-only.

Scope

  • Exactly 11 files stacked on core: make compiled programs the sole session-generation owner #453 (9ff2504…), +1884/-141: seven production/test files and four derived proof/budget attestation files.
  • No client vocabulary, recipe taxonomy, JSON/hex hot path, resource cap, compatibility fallback, WASM export, or controller change.
  • The canonical authored-declaration lowerer into CompiledCoreProgramV1 is intentionally the next stack slice; this PR does not introduce an adapter-authored second Program dialect.

Regression coverage

  • Mixed exact + WCAG evidence in one Program, including joint selection across every admitted case and report-only non-gating behavior.
  • Ready, Stale, Failed-with-remove, Failed-with-hold, exact replay, changed same-revision rejection, malformed width, and duplicate scenario IDs.
  • External-crate RED contract proves package code sees only the concrete owner/session bridge, not generic evaluator or Session machinery.
  • Weak-owner expiry remains a closed package error.
  • Point-support algebra/claims remain unchanged; the verifier renews only exact source binding after the observation/session refactor.
  • The WASM contract pins the measured raw artifact and rejects any uncoordinated budget mutation.

Verification

  • Rust 1.96 full workspace: 714 passed, 6 ignored
  • Integration suites and doctests
  • External package bridge RED test
  • clippy --workspace --all-targets -- -D warnings
  • rustfmt check on the exact production/test overlay
  • point-support independent verifier
  • canonical WASM budget contract test
  • exact-head GitHub CI
  • exact-head Native conformance
  • CodeRabbit

This is a stacked draft. Review only against agent/weak-session-owner; do not merge independently.

@coderabbitai

coderabbitai Bot commented Jul 22, 2026

Copy link
Copy Markdown

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 1d0617b1-e1a8-434f-baad-02d41bf3d14f

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • 🔍 Trigger review
✨ Finishing Touches
🧪 Generate unit tests (beta)
  • Create PR with unit tests
  • Commit unit tests in branch agent/mixed-evaluator-package-bridge

Comment @coderabbitai help to get the list of available commands.

@lemone112
lemone112 force-pushed the agent/mixed-evaluator-package-bridge branch from b8509a3 to 3e8f240 Compare July 22, 2026 23:55
lemone112 added a commit that referenced this pull request Jul 25, 2026
…fact

The exact-length gate still pinned 376830B from
`canonical-authored-program-lowerer` (#456, run 29971399220). This slice
grew the runtime artifact to 376985B and did not carry the re-pin that
every earlier stacked slice performed (#450, #452, #454, #456), so the
gate has rejected this head and every head above it.

The +155B is attributable to this slice alone: runs 30115821523 (#457),
30124467410 (#458), 30125634830 (#459), 30129537515 (#460) and
30136346868 (#461) all measure exactly 376985B, so #458-#461 contribute
zero bytes to the artifact and were failing only on the inherited pin.

The new measurement is the CI run for this exact head
(39ee0a7), not a local build: the
canonical platform is linux-x64 and a local arm64 build only produces a
DIAGNOSTIC result. The budget file's own SHA-256 is re-pinned in the
checker so the drift gate keeps rejecting unattributed edits.

Co-Authored-By: Claude <noreply@anthropic.com>
lemone112 added a commit that referenced this pull request Jul 26, 2026
* feat(core): compile generic point-render sessions

* fix(core): satisfy format and all-target lint gates

* perf(core): recycle point-render session buffers

* style(core): format reusable session buffers

* feat(lcs): execute sealed sRGB8 tristimulus derivation

* style(lcs): format sealed tristimulus slice

* feat(lcs): derive versioned appearance views

* style(lcs): canonicalize Oklab coefficient

* feat(core): expose typed terminal program path

* fix(core): satisfy terminal Program quality gates

* feat(lcs): bind private output projection registry

* style(core): canonicalize release module order

* fix(core): scope private registry lint firewall

* fix(core): harden terminal Program admission

* style(core): match pinned Rust formatter

* feat(core): bind F0 release descriptors

* perf(core): add linear canonical surface ingestion

* refactor(core): unify terminal physical identities

* perf(core): prebind terminal evaluation slots

* perf(core): make present ingestion lazy and borrowed

* refactor(core): remove per-port surface mutation

* feat(core): evaluate typed constraints and emit terminal paints

* feat(core): bind observation groups to runtime streams

* refactor(core): share one encoded point Paint value

* refactor(core): share revision-bound observation state

- admit correlated scenario sets once behind shared immutable backing
- make joint selection domain-safe, linear, and ownership-retryable
- refresh source-bound release proof gates

* fix(ci): bind V2a budget to measured artifact

- keep allocation tests ownership-preserving without a large Result closure
- pin the zero-headroom WASM ratchet to the reproducible V2a artifact

* refactor(core): remove superseded Program runtime

Keep only the private Program compiler/lowering payload for the direct sole-Session bridge.
Delete the duplicate owner, lifecycle, output materialization, and test evaluator scaffolding.
Pin the hard cut with a negative facade gate and refresh the exact point-support source receipt.

* feat(core): execute compiled plans through sole Session

Replace the point-support-specific lifecycle with one sealed monomorphized Session, attach reusable CompiledProgram epochs through strong ownership, and retain complete case-by-constraint reports. Bind every decision to the exact admitted observation before atomic commit and refresh release proof pins.

* feat(core): bind Program assessments to context-bound LCS

* Hard-delete Pair taxonomy from shipping surfaces

* Compile typed finite targets into Program sessions

* quarantine unsupported Display P3 promises

* core: make compiled programs the sole session-generation owner

* core: close the evaluator union and package session bridge

* core: derive a typed CAM16-UCS occurrence view

* core: lower authored physical declarations into the canonical Program

* core: bind Program content identity

* core: name versioned identity discriminants

* fix(ci): re-pin the runtime WASM budget to this slice's measured artifact

The exact-length gate still pinned 376830B from
`canonical-authored-program-lowerer` (#456, run 29971399220). This slice
grew the runtime artifact to 376985B and did not carry the re-pin that
every earlier stacked slice performed (#450, #452, #454, #456), so the
gate has rejected this head and every head above it.

The +155B is attributable to this slice alone: runs 30115821523 (#457),
30124467410 (#458), 30125634830 (#459), 30129537515 (#460) and
30136346868 (#461) all measure exactly 376985B, so #458-#461 contribute
zero bytes to the artifact and were failing only on the inherited pin.

The new measurement is the CI run for this exact head
(39ee0a7), not a local build: the
canonical platform is linux-x64 and a local arm64 build only produces a
DIAGNOSTIC result. The budget file's own SHA-256 is re-pinned in the
checker so the drift gate keeps rejecting unattributed edits.

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(ci): re-bind the point-support source capsule to this slice's cone

This slice moves files inside the point-support semantic cone, so the capsule
digest and the committed surplus proof move with it. Both are now regenerated in
the same commit that causes the drift, matching the convention the rest of the
stack follows; previously the re-bind was batched at #460, which left #457-#459
fail-closed on their own heads and made the stack unmergeable in order.

Numerical review: every proof field is unchanged. Only the source-binding
identities move -- the file hashes of the cone files this slice edits, the
resulting closure digest, the verifier hash and the rolled-up payload hash.
The surplus mathematics is byte-identical.

Co-Authored-By: Claude <noreply@anthropic.com>

* core: project complete program certificate evidence

* fix(ci): re-bind the point-support source capsule to this slice's cone

This slice moves files inside the point-support semantic cone, so the capsule
digest and the committed surplus proof move with it. Both are now regenerated in
the same commit that causes the drift, matching the convention the rest of the
stack follows; previously the re-bind was batched at #460, which left #457-#459
fail-closed on their own heads and made the stack unmergeable in order.

Numerical review: every proof field is unchanged. Only the source-binding
identities move -- the file hashes of the cone files this slice edits, the
resulting closure digest, the verifier hash and the rolled-up payload hash.
The surplus mathematics is byte-identical.

Co-Authored-By: Claude <noreply@anthropic.com>

* core: bind Program operations to exact owner snapshots

* fix(ci): re-bind the point-support source capsule to this slice's cone

This slice moves files inside the point-support semantic cone, so the capsule
digest and the committed surplus proof move with it. Both are now regenerated in
the same commit that causes the drift, matching the convention the rest of the
stack follows; previously the re-bind was batched at #460, which left #457-#459
fail-closed on their own heads and made the stack unmergeable in order.

Numerical review: every proof field is unchanged. Only the source-binding
identities move -- the file hashes of the cone files this slice edits, the
resulting closure digest, the verifier hash and the rolled-up payload hash.
The surplus mathematics is byte-identical.

Co-Authored-By: Claude <noreply@anthropic.com>

* core: single-own compiled observation schemas

* chore(proof): document Python 3.9 zip invariant

* core: define the public Program API

* core: address Program API review

* fix(ci): re-bind public Program proof capsule

* test: close compile-fail sentinel escape

* core: make joint order non-empty by construction and bound evidence cells

`AdmittedFiniteJointOrderV1` stored one flat tuple slice, so an empty
admitted order was representable and the evaluation path carried a
runtime `state_count == 0 -> InternalInvariant` guard to reject it. The
guard proved nothing about the type; it only re-checked a property the
constructor already enforced.

Split the order into `first + rest`. Non-emptiness becomes structural,
`state_count()` is total, and the `InternalInvariant` branch in
`prepare_program_evaluation_buffers` is deleted rather than left dead.
The `joint_state_count: Option<usize>` parameter disappears with it:
cell counts are now derived from the epoch itself.

The same pass stops reserving an exhaustive-conflict buffer that no
constraint can ever fill. `can_conflict` is false when every compiled
constraint is report-only, so a report-only program no longer reserves
`cases x constraints x states` cells and no longer reports
`ResourceExhausted` for a conflict it cannot produce.

`OwnerV1::evidence_cell_bounds` exposes the same arithmetic as a pure
preflight, with `EvidenceBoundsErrorV1::CardinalityOverflow` as the only
closed failure. It creates no Session and mutates no state.

Verified locally on the CI-pinned toolchain: full workspace tests green,
`cargo fmt --all --check` and `cargo clippy --workspace --all-targets --
-D warnings` clean.

Co-Authored-By: Claude <noreply@anthropic.com>

* docs: explain packed proof invariants

* core: fail closed without current Program evidence

* core: revoke outputs on unknown session handoff

* core: keep incomplete Program crate-private

* perf: bind smaller private Program wasm

* core: make staged Program unexportable

* ci: explain rustdoc surface mismatch

---------

Co-authored-by: Claude <noreply@anthropic.com>
@lemone112

Copy link
Copy Markdown
Collaborator Author

Superseded by the reviewed cumulative squash merge #465 (24fd1f4). The lower stacked branch was intentionally not merged on its own because its intermediate head was not the safe terminal public boundary.

@lemone112 lemone112 closed this Jul 26, 2026
@lemone112
lemone112 deleted the agent/mixed-evaluator-package-bridge branch July 27, 2026 00:13
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant