Trust/Payments layer: verifiable capsule receipts (capsule-emit-nanda) — rebuild addressing #177

Description

The pitch.

# Trust/Payments layer: verifiable capsule receipts (`capsule-emit-nanda`)

> cc @Skyrider3 @dhve — this reopens the trust-layer work from #54 → #177, which
> you closed as surface-level. You were right, and the clean review was appreciated.
> It's been rebuilt so the structural work is actually done: each of your four points
> is addressed below, verified against this repo's real `ci.yml` (plain `uv sync` →
> `ruff` / `pyright` strict / `pytest`), and the anchoring validator is scoped to its
> own scenario type so it never touches the stock `receipt_reputation` scenario. Grateful
> for another look when you have a moment.

Re-opens the work from #54 → #177 (closed by @Skyrider3), rebuilt to address every
point raised. This adds an optional **Trust** layer plugin (`capsule_emit`) and a
**Payments** adapter backed by [capsule-emit](https://pypi.org/project/capsule-emit/):
each receipt is sealed into a verifiable JSONL **capsule ledger**, and reputation is
scored only from receipts that still verify against that sealed ledger (a tamper-evident
"Gate 3" that `agent_receipts` cannot provide because it has no ledger reference).

## Every rejection point from #177, addressed

**1. "The plugin is never installed/registered in CI."**
`capsule-emit-nanda` is now a declared **root dependency** (not just a `[tool.uv.workspace]`
member), so the plain `uv sync` that CI runs installs it and registers the
`nest.plugins.trust` entry point. Verified clean-room: `uv sync` → entry point
`['capsule_emit']` → `nest run scenarios/receipt_reputation_capsule.yaml` → ledger written.

**2. "The adversarial validator never runs — nothing is registered in `VALIDATORS`."**
Registered `validate_receipt_reputation_anchored` in `nest_core.validators.VALIDATORS`
under a **new `receipt_reputation_capsule` scenario type** (task mechanics aliased to the
`receipt_reputation` factory; validator set = ring-severance + honest-confidence + anchoring),
so `validate_trace` now enforces it. It is a real trace validator, not a pytest test: every
**signature-valid** receipt observed on the wire must be **anchored** — its independently
recomputed RFC 8785 (JCS) content digest must appear as a sealed `agent_input_digest` in the
capsule ledger the run produced. It verifies the issuer Ed25519 signature too, so "valid
receipt" means signature-valid, not merely well-formed. Proven:
- Capsule scenario, `trust: capsule_emit` → **PASS** — "all 29 receipts anchored in capsule ledger".
- Same scenario, `trust: agent_receipts` (no ledger) → **FAIL** — the reference plugin genuinely
  fails our stronger validator (this is the discrimination dhve asked for, not a parity demo).
- Tamper-after-seal → **FAIL** (digest no longer matches).
- The **stock `receipt_reputation` scenario is untouched** — the anchoring validator is scoped
  to its own type, so we don't break any existing scenario.
The rig now grades the thing being submitted.

**3. "The pyright gate was excluded, not passed (`exclude = ["examples"]`)."**
Removed the exclude; `examples/capsule-emit` is in pyright's checked set and `testpaths`.
All 69 strict errors fixed **by fixing the types** — PEP 561 stub-only packages for the
inline-typed-but-un-`py.typed` `capsule_emit` / `agent_action_capsule`, local
reimplementation of the stock scoring helpers (so there is no cross-module private access),
explicit annotations. **No `exclude`, no `# type: ignore`, no `# pyright: ignore`, no
`reportX = false`.** `uv run pyright` → **0 errors** (full repo).

**4. "`anchor=True` default → live POST; ledger ordering nondeterministic."**
`anchor=False` by default — the graded run does no network I/O — and the ledger write is
synchronous. The graded **trace is byte-identical across runs**. (Anchoring is one flag
away for real adopters: `anchor=True` posts a digest-only record to the free public
transparency log `https://anchor.agentactioncapsule.org/v1/digest` with zero further config;
it stays off for the deterministic benchmark.)

## Also harde

…