fix(gateway): config changes no longer exhaust session discovery with GitHub publication refreshes (#195679)

Root cause: src/gateway/github-publication-discovery.ts:135-146 (main e08aa74ea39e) `refreshAll` schedules a preparation for every remembered session (up to 256) at `delayMs: 0`, and the scheduler starts them all at once. Each preparation reads the session-store inventory on the one-worker, 128-pending "Session target discovery" lane, so one config change fills that lane, refuses user reads with `worker task capacity reached`, and leaves the refused sessions' publication facts unavailable.

Fixes #195649

Thanks @609NFT for the report and the production numbers, and @ElNeMeR70 for the serialization fix this builds on.

## What Problem This Solves

Fixes: user turns and session reads fail with `worker task capacity reached`, and GitHub publication becomes unavailable for many sessions, right after an agent add or other config change on a Gateway with more than about 128 recently active GitHub-bound sessions.

## User Impact

A config change (agent add, model switch) or a GitHub identity change no longer refuses user work on busy Gateways. Publication facts refresh in the background one session at a time, as before at the same ids and times, and all of them come back.

## Why This Change Was Made

Discovery now prepares one session at a time, per discovery instance (@ElNeMeR70's commit). On top of that, every write to the discovery cache goes through one `store()` owner. When the session being prepared is replaced by demand, its binding is removed, or it is evicted from the cache, the queue moves on right away. The stale preparation still finishes on its own and its result is never published. In the first commit only some replacement paths released the queue, so a demand replacement (a new session id for the same key) left the next session waiting behind obsolete work. No new limits or config; worker admission is unchanged.

## Evidence

Gateway run from source with local model and GitHub API fixtures, 256 managed-worktree sessions with publication facts, an agent add through `config.patch` while user turns (`chat.send`) and session opens (`chat.history`) arrive every 10 ms. Full observed output: https://github.com/openclaw/openclaw/pull/195679#issuecomment-6986684891

- main e08aa74ea39e: one agent add left 131 of 256 publication facts unavailable (refused by the 128-pending lane); in another run 19 `chat.history` requests failed with `UNAVAILABLE` `WorkerTaskError: worker task capacity reached: code=overloaded`.
- This head: 0 capacity refusals, 150/150 turns and 150/150 session opens succeed, all 256 facts prepared again (39 s).
- Controls on this head: one remembered session refreshes in 0.76 s as before; a second agent add 1.1 s into a wave replaces it cleanly (all 256 prepared at 47.8 s, no second full wave, no refusals).
- Regression test fails on main (200 concurrent preparations) and on the first commit (queue stuck behind the replaced session), passes here; `src/gateway/github-publication-discovery.test.ts` 12/12.

Merge danger: two-way door. Process-local scheduling only; no stored state, config or protocol changes, and a revert restores the concurrent refresh. Blast radius: GitHub publication discovery on the Gateway (when publication facts refresh after config, identity or profile changes).

Co-authored-by: Ayaan Zaidi <hi@obviy.us>