Harden the cross-company incident response effect boundary

[bordumb]

What changed

• establishes an opaque, native-owned Python authorization waist and brings the Python workflow surface up to the required Rust semantics
• adds matching TypeScript and Python application gateways for ordered plans, durable reservation, credential acquisition, provider execution, outcomes, and signed receipts
• expands the Rust receipt model and exposes canonical preparation, attestation, and verification through Python and WASM
• adds Rust-generated differential fixtures covering canonical commands, commitments, receipt identifiers, and signing preimages
• hardens the cross-company incident demo so the trusted Python service owns authorization and the browser remains only a proposer and observer
• authenticates Northstar through authorization code + PKCE with locally verified ES256 tokens and EdgeShield through an Ed25519 signature over the exact transaction
• reserves execution state before credentials or provider entry, blocks replay and concurrent ownership, preserves unknown outcomes, and requires explicit reconciliation
• carries exact cache command bytes over real Iroh without treating delivery as authorization
• updates the executable specification, threat model, architecture, feature evidence, and README happy-path/failure-path diagram

Why

The original demo illustrated the product idea but did not prove the actual effect boundary strongly enough. This change makes the claim executable: two organizations can approve the same exact, short-lived, ordered authority without sharing an identity provider, root key, or provider credential, and an agent can execute only the native-authorized commands after durable reservation.

It also closes TypeScript/Python parity gaps by keeping canonicalization, authorization, lifecycle, and receipt meaning in Rust while exposing idiomatic, type-driven ports in each SDK.

Security and product impact

• arbitrary TypeScript, Python, browser, HTTPS, or Iroh inputs cannot mint an effect-capable command
• mutation, widening, expiry, approval substitution, replay, concurrency, and wrong-order execution fail before provider entry
• provider credentials are acquired only after durable reservation
• successful delivery never becomes authorization
• ambiguous remote effects remain outcome-unknown, cannot be retried blindly, and require explicit reconciliation
• decision and execution receipts are canonical Rust-owned signed artifacts
• the implementation intentionally carries no prelaunch compatibility aliases, migration paths, or deprecated receipt formats

Validation

• cargo test -p auths-receipts — 6 passed
• cargo check --manifest-path bindings/python/Cargo.toml
• cargo check -p auths-proof-wasm
• Python SDK suite — 73 passed against the rebuilt native extension
• TypeScript full suite — contract checks, 87 integration/adversarial cases, 25 unit cases, package installs, examples, public API, capability metadata, and runtime contract
• cross-company incident demo local suite — Python service tests, live integration, browser controls, real Iroh delivery, concurrency, replay, ambiguity, reconciliation, and EdgeShield Rust tests
• architecture policy and dependency snapshots
• semantic freeze
• binding semantics
• public naming
• core boundary
• specification/registry synchronization