Add M2a plan: thirteen tasks, tests, fake server and recordings

The tasks build the inference path: emsha-backed SHA-256, inferproxy,
config, a hand-written HTTP and SSE client, request building, delta
assembly, the chat state machine, the thinking cap, the slot gate with
retry, the startup self-test and on-device verification.

Everything the tasks copy in was checked against a private reference
implementation: the gate passes after each task in order, the timing
tests pass repeatedly under CPU load, and the reference passes the
self-test and all four device checks on straylight. Expected results
for the recorded streams were derived by a separate script.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

docs/plans/M2a/13-verify-device.md

@@ -0,0 +1,76 @@
+# M2a task 13: verification on the real server
+
+**Branch:** `m2a` (run `git switch m2a`; `git status --short` must be empty, otherwise stop)
+**Commit subject:** `Add on-device verification for the inference path`
+
+## Goal
+
+Run the whole inference path against the real server on straylight and show that it holds: the
+self-test, a capped thinking block, a request that survives its connection being killed, and a
+second turn that reuses the cache of a first turn that contained thinking.
+
+You write no library code in this task. You put two given files in place, run the checks, and
+report what happened. **If a check fails, that is a finding. Do not change a test or a limit to
+make it pass.**
+
+## Context
+
+`make verify-device` builds the workspace, then runs the tests in `crates/loopd/tests/device.rs`
+that are marked `#[ignore]`, one at a time. Each starts its own `inferproxy` pointed at
+`straylight:11434` and talks to the model through it, on slot 0. The "survives" test kills and
+restarts **its own `inferproxy`** in the middle of an answer. Nothing here touches the real
+`llama-server` process.
+
+## Files
+
+- Copy: `crates/loopd/tests/device.rs`, `Makefile` (replaces the old one; only the
+  `verify-device` target differs)
+- Modify: `docs/implementer-log.md`
+
+## Steps
+
+- [ ] **1. Copy.**
+
+```sh
+git switch m2a
+cp docs/plans/M2a/files/crates/loopd/tests/device.rs crates/loopd/tests/
+cp docs/plans/M2a/files/Makefile Makefile
+git diff --stat Makefile
+```
+
+- [ ] **2. Check that the ordinary gate ignores the new tests.** `make gate`. Expected last line:
+  `gate: ok`, and in the output for `device`: `0 passed; 0 failed; 4 ignored`.
+
+- [ ] **3. Check that the server can be reached.** `curl -s http://straylight:11434/health`.
+  Expected: `{"status":"ok"}`. If not, stop and report: this task cannot be done without it.
+
+- [ ] **4. Run the device checks.** `make verify-device`. Expected: `4 passed`, then
+  `verify-device: ok`, in well under two minutes. The first run may take longer if the model has to
+  be loaded.
+
+- [ ] **5. If a check failed,** run it alone to see its output, for example:
+
+```sh
+BOXMAKER_INFERPROXY=$PWD/target/debug/inferproxy BOXMAKER_UPSTREAM=straylight:11434 \
+  cargo test -p loopd --test device -- --ignored --test-threads=1 --nocapture a_capped_thinking
+```
+
+  Put the test's name and its message in your log row with status `stopped`. A failure here is
+  either a real defect in an earlier task or a change on the server; the reviewer decides which.
+
+- [ ] **6. Log and commit.** Put the four results and the total time in your log row.
+
+```sh
+git add Makefile crates/loopd/tests/device.rs docs/implementer-log.md
+git commit
+```
+
+## Done when
+
+- `make gate` prints `gate: ok` and `make verify-device` prints `verify-device: ok`.
+- `cmp Makefile docs/plans/M2a/files/Makefile` prints nothing.
+- `git log --oneline master..m2a` shows one commit per task.
+
+## Stop and report if
+
+- The server cannot be reached, or any device check fails twice.