SynthWorld 0.13.0

Added

• A deterministic enterprise identity and access surface, with authorization
  evaluated against standards-shaped projections (#7, #27). SynthWorld now
  generates a fixed enterprise identity/access universe — organisations, units,
  populations, principals, accounts, groups, roles, permissions, entitlements,
  and opaque authorization targets — and evaluates access against it through
  three bounded oracles: a directory RBAC oracle with role and group closure,
  a bounded ABAC oracle over declared attributes, and a bounded ReBAC
  oracle over relationship tuples. Access derivation is an implementation detail
  of the oracle rather than a published topology, so no artifact describes the
  model as an identity topology.
  New modules: enterprise, enterprise/rbac, enterprise/abac,
  enterprise/rebac, enterprise/authorization, enterprise/conformance,
  enterprise/identity_fabric, agentic/enterprise.
• Vendor-neutral projections so a real authorization engine can consume the
  world without a bespoke adapter. enterprise/projections/ emits an RFC
  7643-style SCIM user/group projection, an OpenFGA authorization model
  and relationship tuples for the bounded ReBAC subset, and an AuthZEN 1.0
  request projection with per-field provenance. Each projection carries an
  explicit mapping profile recording what it can and cannot represent, so a
  projection gap is declared rather than silently lossy. A Shared Signals/CAEP
  mapping profile is declared, but temporal event emission is deliberately
  deferred and no SET envelope is constructed yet.
• Two smoke benchmarks and a graded assurance ladder. An enterprise identity
  fabric smoke benchmark and an enterprise agentic smoke benchmark publish public
  inputs and evaluator truth under enterprise-identity-access-contract/; a
  contextual access benchmark profile publishes under
  contextual-access-contract/. continuous_assurance adds smoke,
  standard, longitudinal, and held_out tiers governing assurance cadence.
  Note these are assurance tiers, not generated-world scale tiers — the
  enterprise_agentic scale ladder #27 asks for remains open at smoke only.
• An executable agent-authority run protocol. agent_authority and
  assurance add a staged run protocol with signed execution receipts, component
  provenance, and evidence claims, so an evaluation run is reconstructible from
  its receipt rather than trusted on assertion.
• The ambiguity v2 pack gets its difficulty from a computed error floor, not a
  codebook (#80). The v1-style surfaces encoded each identity index in cleartext, so a
  ~30-line normaliser recovered every relation and scored 1.0000; two successor designs
  fell to pool inversion. The fix follows the reviewed plan: each kind draws a base from
  a pool of confusable clusters (Sorensen/Sorenson/Soerensen, a transposed
  phone, a swapped day/month), EQUAL/NEAR share the base while FAR redraws from a
  stationary mixture, and every value passes through one structured-noise operator
  applied identically under every relation. Identity recovery stays free and expected;
  the relation is carried by overlapping distance distributions. The pack publishes its
  genie floor — the Bayes error of the generator, estimated with a stated N and
  Wilson interval and keyed to a digest of every decision-relevant constant — plus the
  enumerated channel invariants (kernel stationarity, an identical one-value marginal
  under every relation, a per-base sibling-landing mass gate, form bijectivity and
  constant cross-form distance, and an artifact-factorization check) and a gated
  technique premium, so real resolution technique is rewarded rather than anti-taught.
  New modules: ambiguity_evidence, ambiguity_surfaces, ambiguity_channel,
  ambiguity_floor, with v2 serialization/metrics/baselines support and
  examples/compute_ambiguity_floor.py.
• Agent-authority and contextual-access receipt builders now seal honest
  failed-run receipts: a failed product execution produces a manifest with
  execution_status=failed and evaluation_status=not_evaluated that binds
  only the product-stage artifacts and never loads evaluator truth, so an
  assurance corpus can no longer be structurally biased toward successful runs.
  Receipt validation enforces the paired statuses and the product-only artifact
  inventory, and run manifests whose evidence claim is not supported by the
  systems under test (live-lab claims with reference-only components) are
  rejected. Managed-service provenance in not_exposed observability states
  additionally forbids evidence references, and the contextual execution
  receipt leaves stimulus_digest unset because that lineage executes the
  public input directly.
• Contextual-access receipts now expose the same explicit two-phase live-run
  boundary as agent-authority receipts. External runners may finish and attribute
  the product stage before constructing completion metadata; the finalizer then
  replays the plan, public input, adapter, component inventory, provenance, and
  artifact digests before evaluator truth is loaded. Existing deterministic
  contextual receipt bytes remain unchanged.
• An opt-in disposable agent-authority reference deployment now executes the
  public enterprise-agentic smoke world across isolated Docker networks. It
  produces live observation-v2 receipts covering L01-L06, the exact declared
  L07 baseline/SUT inventory, and measured/unsupported L08 targets, while
  keeping runtime credentials in a destroyed named volume and scanning canary
  and token markers out of receipts, logs, and container metadata. A new
  two-phase receipt finalizer lets live runners record completion metadata only
  after external execution without exposing evaluator truth before product
  output is durably staged.
• Agent-authority observation schema 2.0.0 corrects L06 clock semantics without
  changing the frozen observation-v1 schema. It records one explicit monotonic
  revocation epoch, non-negative acknowledgement offsets, and signed send/completion
  offsets so pre-revocation in-flight requests are representable. Receipt validation
  dispatches v1/v2 observations and binds them to scoring formulas 1.0.0/2.0.0;
  migration guidance forbids guessing offsets from ambiguous v1 rows.
• The 12-case authority-change governance conformance fixture from #73 is now an
  additive frozen benchmark. Its public and evaluator payloads remain physically
  separate; their visibility manifests and exact raw bytes are path-bound by a
  packaged SHA256SUMS, verified by the packaged loader API, regeneration tests,
  and isolated-wheel checks. No existing golden bytes changed.
• The broker-removal pack is scored through the unified evaluator: evaluate_broker_removal
  projects each family's headline ratio into the standard EvaluationReport, the CLI gains
  synthworld evaluate broker, and examples/evaluate_broker_adapter.py is the worked
  Idcognito-style adapter #5 asked for - public timeline in, versioned assessment out,
  scored against regenerated truth. Closes the last acceptance criteria of #5.
• Propagation lag is representable and scored (#65). Downstream copies carry their own
  removal tick (None never goes), a new slow_propagation case has copies that catch up
  late rather than never, a submission can predict the completion tick, and
  propagation_lag reports mean absolute error with support. The credulous baseline now
  predicts completion at confirmation - "done means done everywhere" - and eats a 14-tick
  error on exactly the case built to price that claim; the example adapter's modest grace
  period cuts it to 4.

Changed

• Ambiguity grammar 2.0.0, v2 schema 2.1.0. render_relation/render_value
  delegate to the structured-noise channel; the old _SPACE/_surface codebook is gone.
  Relation.EQUAL no longer means "byte-identical" — it is one value transcribed once
  per record, rendered identically only with probability sigma — and the charter
  docstrings that claimed otherwise are rewritten. The v1 pack and its frozen artifacts
  are untouched and stay byte-identical.
• display_name renders family, given (#86). The two name kinds are scored as
  separate evidence, so the boundary between them must be readable off the value; the
  old given family lost it whenever a pool entry carried a space. No rendered name
  contains ", ", so the split is unambiguous.
• Temporal schema 1.2.0. ListingTruth.downstream_refs (bare strings) becomes
  downstream_copies with per-copy removal ticks, and a recorded reappearance must now
  coincide with a published LISTING_REAPPEARED event - truth that disagrees with the
  public timeline is refused as corrupt input. Deliberately asymmetric with removal, which
  stays unpinned because a confirmation is the broker's claim and the phantom case exists
  to show the claim can be false. BrokerAssessment moves to 1.1.0 for the new
  prediction field; propagation state is now read as of the assessed tick, so a slowly
  propagating deletion no longer scores identically to one that never propagates.
• DenominatedMetric moved to synthworld.models, below the evaluation/partition import
  cycle it was about to create; ambiguity_partition re-exports it unchanged.

What's Changed

• Close the broker pack: lag scored, evaluator unified, adapter worked (#5, #65) by @bluntmachetti in #95
• Give the ambiguity v2 pack a computed error floor, not a codebook (#80) by @bluntmachetti in #96
• Add the enterprise authorization testbed: identity fabric, contextual access, and governance benchmarks (PR1-PR13) by @bluntmachetti in #97
• Prepare SynthWorld 0.13.0: document the enterprise authorization surface by @bluntmachetti in #99

Full Changelog: v0.12.0...v0.13.0