Bump version to 0.2.0:58 (agent-payment-connect)

Ships the agent-delegable BTCPay connect gate: a scoped key with the
payment_providers:write a-la-carte scope can connect a non-mainnet BTCPay
provider on a sandbox daemon, fail-closed; master/mainnet/production stay
master-only. Migrations 0024-0025 additive; openapi documents the BTCPay
paths. api suite 65. No SDK change.

startos/versions/v0.2.0.ts

@@ -39,6 +39,8 @@ const RELEASE_NOTES = [
 // in RELEASE_NOTES above (the milestone). Subsequent revisions
 // append here.
 const ROUTINE_NOTES = [
+  '0.2.0:58 — **Agent-delegable BTCPay connect, gated to sandbox + non-mainnet.** Makes Keysat fully agent-operable for *dev/test setup*: an operator can hand an agent a scoped key that connects a BTCPay payment provider over the API — no master key, no browser click — but only on a sandbox daemon and only for a non-mainnet (regtest/testnet/signet) store. On a production daemon, or for a mainnet store, connecting a provider stays master-only, and disconnect is always master-only. The reasoning: a credential that can repoint where settlement lands is a fund-redirection key, so the capability is deliberately narrow and fails closed. **Gated in three layers:** (1) a daemon-level `KEYSAT_SANDBOX_MODE` flag, read at boot and never settable via any API, is the outer gate — scoped connect is disabled entirely on a production box; (2) `payment_providers:write` is an à-la-carte per-key scope that belongs to no role (not even full-admin), granted explicitly when an operator mints a key; (3) at OAuth-callback time the daemon resolves the target store\'s Bitcoin network from its on-chain receive address and refuses anything not provably non-mainnet, failing closed to mainnet on any ambiguity (no on-chain wallet, unreachable BTCPay, unrecognized address) — it denies rather than guesses. Migrations 0024 (`scoped_api_keys.extra_scopes`) and 0025 (`btcpay_authorize_state.scoped_initiator` + actor hash, to carry the initiator across the browser round-trip) are additive — straight drop-in over :57. The served OpenAPI spec now documents the BTCPay connect/callback/status/disconnect paths and the key-creation `scopes` field, and `/v1/admin/tier` surfaces a read-only `sandbox` flag. Also hardened: the GET authorize-callback now returns the real HTTP status on a denied connect (was a misleading 200 with an error page). Validated end-to-end against a live regtest BTCPay; the docs-onboarding harness (a fresh agent integrating from the published docs alone) converged completed-clean on the full buyer-pays journey. Daemon api test suite is at 65, up from 57. Zaprite connect stays master-only. No SDK change.',
+  '',
   '0.2.0:57 —**New `merchant-onboard` scoped-API-key role for least-privilege self-serve onboarding.** A fifth scoped-key role sits between `license-issuer` and `full-admin`, granting read access plus `products:write` + `policies:write` + `licenses:write` — the minimum a merchant (or an integrating agent) needs to stand up a catalog end-to-end over the API: create a product, define its policies/tiers, and issue licenses against them, all without holding the master key. The catalog write *scopes* already existed and were enforced on the endpoints since :55; only a role that expands to them was missing, so this is a `Role`-variant addition, not a scope-model change. `Role::grants` matches the write scopes explicitly (never by `:write` suffix), so the role can never widen into settings / payment-provider / merchant-profile / webhook writes, and every master-only operation (signing-key rotation, payment connect, web-admin password, API-key management, server settings, per-license tier change, DB introspection) stays behind `require_admin` and is structurally unreachable from any scoped key. Existing Creator-tier caps still bound it (5 products / 5 policies per product / 10 active codes). **Caveat:** the role covers catalog setup + manual license issuance fully, but connecting a BTCPay/Zaprite payment provider stays master-only by design, so the buyer-paid purchase flow still needs a one-time operator step. Migration 0023 rebuilds `scoped_api_keys` to widen the role CHECK constraint (SQLite can\'t alter a CHECK in place; the table has no foreign keys, so it\'s a plain copy/drop/rename) — additive, a straight drop-in over :56. Daemon api test suite is at 57, up from 56. No SDK change.',
   '',
   '0.2.0:56 — **Product→merchant-profile write path — multi-profile is now functional end-to-end.** The multi-profile *resolver* has been complete since :52, but products had no way to be *assigned* to a profile, so every product stuck to the auto-created default profile. This cut wires the missing write half. `Product.merchant_profile_id` now threads through all four product SELECTs + `row_to_product`; a new `repo::set_product_merchant_profile` validates the target profile exists first (returns a clean 404 rather than a raw FK 500); it is threaded through `CreateProductReq` (applied as a post-write step) and `UpdateProductReq` (double-`Option` semantics, where `Some(None)` clears a product back to the default profile). The admin SPA shows a merchant-profile `<select>` on the product form only when more than one profile exists, so single-profile operators see no change. No schema migration (highest is still 0022) — straight drop-in over :55. No SDK change.',
@@ -530,7 +532,7 @@ const ROUTINE_NOTES = [
 ].join('\n\n')
 
 export const v0_2_0 = VersionInfo.of({
-  version: '0.2.0:57',
+  version: '0.2.0:58',
   releaseNotes: { en_US: ROUTINE_NOTES },
   // No on-disk transformation needed — v0.2.0:0 is a label change.
   // SQLite-level migrations live separately under