docs(seam): delete the false FLUID alignment claim; gate the vendored copies

[fas89]

FLUX claimed — in docs/reconciliation.md and in the published schema's own $defs/agentPolicy description — that its agent-policy vocabulary was "aligned with FLUID's" and that "seam-crossing policy comparison is field-to-field".

Neither is true.

Measured against fluid-schema-0.7.5.json

	FLUX agentPolicy	FLUID exposes[].policy.agentPolicy
members	5	13
token budget	tokenBudget (integer)	maxTokensPerRequest + maxTokensPerDay
purposeLimitation	boolean	string
use-case values	free strings (minLength: 1)	12-value enum
FLUID-only	—	deniedModels, canReason, canStore, retentionPolicy, auditRequired, tags, labels

Three members line up by name and type: allowedModels, allowedUseCases, deniedUseCases.

The use-case row is the one that bites

FLUX accepts any non-empty string; FLUID accepts twelve specific values. So a ConsentProfile can deny a use case that no FLUID contract is able to express, and the "at least as strict" seam rule then has nothing to bind against on the FLUID side.

That is not hypothetical. Of the eleven distinct use-case values in FLUX's own 34 enforcement vectors — advertising, persona_variation, journey_synthesis, pricing_optimisation, exfiltrate-pii and the rest — none can be written into a FLUID contract. The comparison the old wording described has never had a working example in this repository's own test suite.

What the validator does enforce is real and still regression-tested: within FLUX, a simulation's declared policy is checked against its gating profile. The claim that this reached across the seam was the false part.

What changed

• docs/reconciliation.md now carries the divergence table above, says plainly what does and does not cross the seam, and records two ways to resolve it — FLUX $refs FLUID's published agentPolicy $id (cleanest, breaking, cheapest now while FLUX has no external implementers), or FLUX keeps its own vocabulary and the table stands as the permanent honest statement. That is an open decision, recorded as one rather than pre-empted.
• The same false sentence is removed from $defs/agentPolicy's description in flux-schema-0.5.0.json and the latest alias, with the mirror under flux_spec/data/ regenerated by scripts/sync-package-data.py.

Fixed in place rather than deferred to a version bump: 0.5.0 is untagged and flux-spec is not on PyPI (404). Publishing first would have shipped a citable falsehood to a package index.

Also: a gate for the vendored FLUID schemas

.github/workflows/fluid-vendor-sync.yml compares vendor/fluid/*.json against what FLUID publishes.

Correction to an earlier internal assessment: those copies are not drifted. I checked all three (0.7.3, 0.7.4, 0.7.5) canonically and they are byte-identical to the published schemas. The gap was never drift — it was that nothing enforced the absence of drift, and an unguarded cache ages. The job mirrors FLUID's own schema-sync.yml, compares canonical JSON so formatting churn can't fail the build, and runs weekly because FLUID can publish without FLUX touching anything.

The rule behind both halves of this PR: a cross-boundary claim is only true if the CI of the side making the claim proves it.

Tested

FLUX's own checks, with jsonschema, pyyaml and rfc8785 installed:

scripts/validate.py         exit 0
tests/test_regression.py    ok — 132 regression checks green
sync-package-data.py        mirror regenerated
latest alias                still byte-identical to 0.5.0

(Note for anyone reproducing: without rfc8785 installed, validate.py and the regression suite both exit 1 with a dependency message on a clean checkout — that failure is environmental, not a regression.)

🤖 Generated with Claude Code