[rwkeyes]

Problem and result

conv_route_batch_after in the run config was ignored. serve/server.py read a.conv_route_batch_after — the
--conv-route-batch-after flag, default 4 — while bin/strata-b70-serve.sh writes the key into the config JSON and
starts the server with only --engine, --config and --port. A deployment that set 1 or 2 in the config got 4, and
nothing reported the mismatch.

The fix resolves the value in one place: explicit flag → config key → 4, with malformed values falling back
instead of raising. The flag's argparse default becomes None so "absent" is distinguishable from "explicitly 4".

How it was found

A mixed-traffic sweep set conv_route_batch_after in two configs (4 and 1) and got results identical to 0.02 s on all
eleven steps. Passing --conv-route-batch-after 1 by hand reproduced the 2026-11-09 result immediately — the
conversation's second turn re-prefilled at 2.465 s against 0.587 s. So the config key was inert, not the router.

Validation

Passed:

• python -m pytest serve/test_conv_router.py -q → 12 passed, including the new
  test_the_config_key_is_honoured_when_the_flag_is_absent, which pins: config 2 → 2, config "3" → 3, flag 1 beats
  config 2 → 1, no flag and no key → 4, and 0 / -1 / True / "off" / [1] / {} → 4 rather than an exception.
• import serve.server in the fix worktree → ok, and resolve_batch_after(None, {"conv_route_batch_after": 2}) → 2
  at the real import path, so the new name resolves where the server uses it.
• python -m pytest serve/test_server.py -q → 376 passed, 526 subtests passed in 110 s, on the fix branch in a
  throwaway worktree (the default branch's tree was never touched).
• Source confirms both halves: default=None on the flag, and batch_after=resolve_batch_after(...) at the call site.

Failed:

• -k "conv_route or batch_after" on serve/test_server.py deselects all 376 tests — the wiring has no test there,
  which is why the key could rot unnoticed. The new test lives with the router, where the decision does. (An earlier
  attempt at the same run returned pytest exit 4 because --timeout is not installed in this venv, and exit 5 from the
  filtered run; neither was a test failure.)

Merge note — do not merge into dirigo/0.2.12-verified

That branch keeps src/, include/, sycl/, metal/, serve/, setup.py and CMakeLists.txt byte-identical to
the deployed snapshot so a re-clone of v0.2.12 plus the carry still reproduces the tree hash (DIRIGO.md). Merging this
would break that check. The deployment is not affected today — its launcher config says 4, which is what it gets — so
this should land with the next tree bump.