feat(agents-api): add safe standalone local init

[sam2tom]

Context

Running the standalone Agents API locally currently requires operators to hand-create a caller bearer and six-field keys.json binding before Web or another client can authenticate. That lifecycle belongs to Parsar, not to each UI repository.

Observable outcome: Parsar provides a supported, secure, idempotent initialization command for a standalone local Agents API deployment.

Acceptance criteria

• Add an explicit standalone local-init command owned by services/agents-api.
• Create a canonical nonzero tenant ID, caller principal binding, plaintext caller token file, and digest-only keys.json entry using cryptographically secure randomness.
• Use stable documented default paths compatible with Agents Core Web's server-side proxy convention.
• Refuse to overwrite or silently reuse an existing token, binding, tenant, or device profile unless the operator selects an explicit reviewed mode.
• Create private directories/files with restrictive permissions and verify them after writing.
• Never print the plaintext caller token, daemon credential, authorization header, or provider credential.
• Keep product users/organizations separate: organization/project/subject values are execution-service identities supplied or defaulted explicitly.
• Emit redacted human-readable status and optional machine-readable non-secret metadata for downstream tooling.
• Provide tests for fresh init, repeat invocation, partial state, invalid IDs, permission failure, digest correctness, redaction, and crash-safe writes.
• Document rotation and cleanup semantics without deleting data by default.

Validation

• Run focused Agents API CLI/config tests and the repository's standard Agents API checks.
• Prove no generated secret appears in stdout, stderr, argv, snapshots, or Git fixtures.
• Validate compatibility with the documented Agents Core Web token-file contract.

Non-goals

• Do not add browser or product organization logic.
• Do not provision model/provider credentials.
• Do not start a paid Turn.
• Do not make the root Parsar product stack a dependency of the standalone Agents API.