protocol: reuse-feedback-obs, SPEC 10.1 - feature negotiation, and each observation sent once

[tactino]

Every observation crosses the link twice: once in the feedback that ends a chunk, and again in the next infer as the state the new chunk starts from. The two differ only after a reset. That second copy is the factor of two in E43's cost model (2.8 ms + 2 × observation bytes / bandwidth). This removes it, and adds the negotiation it needs to stay compatible both ways. Stacked on #10 (and #9, #8): merge those first. The three commits after the merge of #8's branch are new.

Negotiation (SPEC 10).

• A server lists the features it implements in metadata's features.
• A client uses a feature only if it is listed there.
• A server that lists one still accepts every client that does not use it.

So a new client works with an old server, and an old client with a new one. protocol_version stays 1. A change that cannot be made this way is the one that would move it to 2.

reuse-feedback-obs (SPEC 10.1). An infer may carry reuse, a bool array with one entry per row:

• A row it marks sends no observation. The server takes it from the obs of that env's last feedback on this connection.
• data is batched to the rows that are not reused.
• A client MUST NOT reuse for an env whose last feedback ended its episode, since that observation was the terminal one. It also MUST NOT reuse for an env with no feedback on this connection yet, which covers every env after a reconnect.
• A server MUST close for a resync on a reuse it cannot honour.

What is in it:

• plugrl_protocol.reuse.ObservationCache: the server's half, one per connection. on_feedback stores each env's row, or marks that its episode ended. complete rebuilds the full observation of an infer and raises ReuseError on a forbidden reuse. on_feedback never raises. A server feeds it every feedback, from clients that never reuse too, so a feedback it cannot split (no text key, say) only makes those envs not reusable.

• plugrl-conformance --features reuse-feedback-obs: offers the feature to a client. The probe checks still see the full observation, so every other clause is graded as before. A client that reuses something the server does not hold fails 10.1. A client that is offered the feature and never uses it gets a note. Without the flag nothing changes, and a client that sends reuse anyway fails.

• plugrl-conformance-server: a new reuse phase, which runs when the server lists the feature. It checks four things:

  • an infer with some rows reused is answered;
  • an infer with every row reused (an empty data batch) is answered;
  • a reuse right after a terminal feedback closes for a resync;
  • a reuse on a fresh connection closes for a resync.

  A server that does not list the feature gets one note.

• Reference implementations: raw_client.py uses the feature when offered (--no-reuse turns it off), and reference_server.py offers it.

• Bugs that show each clause is checked: reuse-unoffered and reuse-after-reset on the client, lenient-reuse on the server.

Checks:

• Local run: 71 passed, 17 of them in tests/test_reuse.py, which unit-tests the cache on its own. ruff 0.14.2 is clean.
• plugrl-server main (835f691) passes the new server checker with no violations and one note, because it does not offer the feature yet.

Before this commit, the branch merges in #8's last commit (the 7.6 refinement), so that #10's branch carries all of #8.

Users:

• plugrl-server#109 offers the feature, using ObservationCache. A reusing client trains the same weights bit for bit.
• plugrl-env-client#16 uses it when offered.
• E46 re-measures E43's ladder next, interleaving v1 and v2 runs.