fix(mcp-server): every knowledge write says what the orchestrator did, and a refused daemon batch is retried - #409

…, and a refused daemon batch is retried

task_1790588967474_xzgt: claude3-x402's re-read of #319 at bd05d52 (room
msg_1790219322472_d0cjwe .. msg_1790219332311_ac825g), and the artifacts
in C:\holo-dev\.scratch\claude3-review-artifacts\319b\.

P2-A. team-routes wrote its mirror before the orchestrator, but no test
pinned that order, and mutants N05/N21 survived all 269 tests. New test:
the stand-in goes silent on the write only, and the team's knowledge is
read while the POST is still waiting. Both rows must be there. It also
checks the new warning line.

P2-B. Four routes and a tool answered a refusal as success, though none
keeps a copy of its own:
- POST /contribute answered 201.
- POST /knowledge/private answered 201.
- POST /knowledge/promote answered 201 (not in the review's list, same
  class).
- DELETE /knowledge/private/:id answered 200.
- The holomesh_contribute tool answered success:true, synced:0.
Each now answers 502 when the orchestrator refused and 503 when it could
not be reached. The routes do this through one helper,
answerUnacceptedWrite in holomesh/utils.ts, which #319's
POST /knowledge now uses too; the tool says success:false. The error names
the status and reason, never the orchestrator's text.

holomesh_publish_tool stays a success, because its manifest really is
published locally. It now says whether the knowledge copy landed.

The daemon marked a batch contributed whatever the orchestrator answered,
so a refused entry was never retried. Both daemon sites now mark nothing
unless every row was taken.

P3, in the client:
- An answer saying success:false with no count is no longer read as the
  whole batch accepted.
- A leading byte-order mark is stripped before JSON.parse.
- An answer over 64 KiB is named as that, not as "without a JSON body".
- The timeout setting is clamped. Node's AbortSignal.timeout throws on
  values above 2^32-1, and every request then read as unreachable.
- The team route logs one line (count and reason, no body) when its rows
  live only in the mirror.
Not done: the test's stand-in port does not retry on EADDRINUSE. The port
is fixed before import (vi.hoisted), and a retry would need a restructure;
the collision odds are about 1 in 20,000 per run.

Verified:
- 10 new tests fail before the fix, and 1 new daemon test fails on the old
  daemon. The ordering test passes on the correct order: it is a pin.
  Controls stay green: accepted writes still answer 201/200, and the tool
  still succeeds.
- The outcome, client and daemon test files: 117/117.
- Each part broken alone: 13 of 13 red, each on its own test. The parts
  are: the mirror order, the helper, each of the four routes, the tool,
  the daemon, success:false, the byte-order mark, the oversized answer,
  the clamp, and the log line.
- tsc 0 errors. The files that were prettier-clean stay clean. The outcome
  test file was already unformatted from #319; it is formatted on its own
  in the next commit.

Stacked on #408 (claude1/orchestrator-client-timeouts): the two changes
share the test file and the client.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>