Open routing foundation: bounded discovery and late-join progress

[qh-work]

Adds an explicit open contact-discovery profile without a common authority or global roster: signed control, bounded two-view routing, leased contact shards, durable revision/revocation/conflict floors, a Python HTTP node/client, and independent native TypeScript control/routing/state kernels.

Late nodes can make progress through full introduction queues. Verified seed reachability survives signed backpressure; verified replacement entries can be introduced without exploratory eviction; exact live peers already proven by challenge do not refill pending capacity. Unknown or changed descriptors and failed probes still require independent challenges. Public rejoin/restart uses the newest known signed seed, and owners can propagate signed contact withdrawal.

Every fourth maintenance lookup refreshes the owner's own region so later bounded announcements reach newly discoverable neighbours. The other three cycles retain random targets. Each cycle still shares 16 RPCs, 1 MiB responses and five seconds across pending challenges, announcements and one lookup. The schedule is explicit in reports. Compact diagnostic serialization preserves complete failure lists within unchanged report limits.

Packages this first slice as 0.28.0-alpha.0, with matching runtime/plugin/TypeScript metadata and download links. Open messaging and the complete TypeScript open Agent/HTTP runtime remain unsupported. The full 0.28.0-alpha.1 consent, encrypted delivery and replica-repair milestone remains ahead.

Validation:

• Functional source dc485334f8ad8629db68ef25c7c618a732f15c6c: three-seed cloud run 34191451681 passed. Seeds17/29/43 each returned 1,000/1,000 healthy and 1,000/1,000 bootstrap-exit lookups, meeting the original independent 99%/97% gates. Complete reports and source bindings were checked; earlier failed runs are retained in the runtime evidence document.
• Open CI 34191411538 passed 84 distinct methods, including native TypeScript vectors. Fifty selected local methods also passed; counts overlap and are not added. Focused regressions reproduced the pre-fix failures and passed after repair.
• Packaging candidate bd2b92cfb22f12648502cce78d2558a6e380fa15: 18 existing packaging/core methods passed in 3.726 seconds; the extracted 56-module client preserved three synthetic records through save/recall, retries, hooks and backup/restore. Other than the VERSION constant, all 215 other Python/TypeScript files remain byte-identical. Dependency metadata changes only its own package version. The scale run remains pinned to dc48533, not relabelled as a new execution.
• Public source and all four archives were reviewed: 709 members, 343 review source files matched the exact Git bytes, and all eight asset checksums matched. No private installation or data was used.
• Current normal CI: 34193349219. Current light CI: 34193349185. Actual 6 Pro implementation and complete-report evidence reviews passed the dc48533 first slice with zero new blockers. The bd2b92c packaging review also passed with zero blockers: 18 methods independently passed plus the actual extracted client package check.

Scale uses real Ed25519 with logical in-process transport, not 100-node HTTP/SQLite, actual AI instances or independent physical failure domains. Bootstrap exit inherits healthy-phase learning. Descriptor renewal, complete native TypeScript operation, first contact/consent, encrypted mailboxes/receipts, repair and larger/hostile/multi-model acceptance remain later gates. Previously published releases are preserved.