config: publishing is the process entry's job

[ch-wan]

Motivation

Publishing the config bags was happening inside constructors — TokenizerManager,
ModelRunner, the encoder server. That makes "is the config published?" depend on which
object a process happened to build first, and it means a process that legitimately builds
none of them (a spawned encoder worker, a benchmark entry) reads unpublished bags. It also
puts a process-wide side effect in an object's __init__, so a test that builds two of
them publishes twice.

Separately, whole-object readbacks (/server_info, the resolved-args dict) walked the
record's fields. Once a decision lives in the declaration stash rather than in the field,
walking fields reports the wrong shape.

What changes

• publish(server_args, role=...) is called by the process entry: the HTTP entry, the
  encoder gRPC entry, the encoder DP and TP workers after spawn, and the one-batch
  benchmark entry.
• Constructors call assert_published(server_args, role=...) instead of publishing. It
  fails closed with the role in the message, so an entry that forgot to publish says so at
  startup instead of returning a stale or empty namespace.
• Whole-object readbacks go through resolution_projection(server_args) /
  ServerArgs.resolved_dict(), which report the declaration result for every leaf,
  expanding nested dataclasses and namedtuples the way asdict does.
• test_publish_precedes_bag_reads.py walks each known entry and asserts publishing
  precedes the first bag read.

Nothing here changes what any field holds: resolution still writes as it declares, and the
end-of-resolution replay is untouched. Both go away together in the last PR of the stack.

How to verify

export PYTHONPATH=$PWD/python
python -m pytest -q test/registered/unit/test_publish_precedes_bag_reads.py
python -m pytest -q test/registered/unit/test_runtime_context.py
python -m pytest -q test/registered/unit/server_args/test_resolution_declarations.py

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #32965569690
Latest PR Test (Extra): ✅ Run #32966000247
Latest PR Test (AMD ROCm 7.2): ❌ Run #32965569687