Add compact summary output format to list tools (bounded, docker ps-style)

[mecworks]

Summary

list_containers, list_images, and list_volumes returned the full
inspect-level shape for every object. That's fine for a single object, but
unusable for a listing: the payload is dominated by data a docker ps-style
answer never needs, and it's unbounded — it grows with what each
container's healthcheck prints.

On a 22-container host, a single list_containers call returned ~103 KB,
roughly 90% of which was health-check log output. For an LLM-facing tool, that
is the classic "inspect-everything" antipattern: the server can't know what the
model will need, so it should default to a minimal view and let the caller opt
into detail.

Root cause

docker_to_dict() (shared by every tool) embeds, for each container:

• state = the raw inspect State, which includes
  Health.Log — Docker retains the last 5 healthcheck outputs and never
  bounds their size. A healthcheck that curls a web page (or prints a JSON
  status dump) therefore adds tens of KB per container, per listing.
• image = a nested docker_to_dict(image) — repo tags, digests and OCI
  labels repeated for every row.
• mounts = full mount definitions (source paths, driver, propagation).

remote-shell-mcp (a different Docker MCP) solved this by making list and
inspect separate verbs with a fixed 6-field row type. This PR does the same
thing within a single format parameter.

What this changes

Adds a format parameter — "summary" | "full", default "summary" —
to the three list tools.

Summary returns a compact docker ps-style row:

name, short_id, status,
state.{Status, Running, Restarting, Dead, ExitCode, StartedAt},
image            (string ref, no nested object),
mounts           (trimmed to name + type),
health.{Status, FailingStreak}   (no Log)

plus a slim image/volume row (no labels/digests).

Full keeps the existing inspect shape for debugging. In both modes,
every Health.Log output entry is capped at 500 characters, so even a full
listing is bounded regardless of what a healthcheck prints.

Single-object tools (run_container, stop_container, …) are unchanged:
docker_to_dict() still defaults to format="full".

Behavior change (please read)

The default for the three list tools flips from inspect-level output to
the compact shape. Any existing client that parses the old fields from a
list_* call will need to pass format: "full" (or read the summary fields).
This is the intended fix — the old default was the bug — but it is a breaking
change for anyone scripting against the old list output.

Example output (real 22-container host, before → after)

Call	Before	After	Reduction
list_containers (all 22)	102,965 chars	11,048 chars	89%
one web-app container row	24,013 chars	326 chars	99%
another web-app row	19,629 chars	323 chars	98%
a status-dumping proxy row	11,201 chars	424 chars	96%
list_images (27)	24,119 chars	4,666 chars	81%
list_volumes (13)	6,210 chars	3,189 chars	49%

A summary row:

{
  "name": "webapp-a",
  "short_id": "11aef40ef8cb",
  "status": "running",
  "state": {
    "Status": "running", "Running": true, "Restarting": false,
    "Dead": false, "ExitCode": 0, "StartedAt": "2026-09-15T15:55:05Z"
  },
  "image": "exampleorg/webapp-a:latest",
  "mounts": [],
  "health": { "Status": "healthy", "FailingStreak": 0 }
}

(In full mode, Health.Log entries longer than 500 chars are truncated with
a … [truncated] marker.)

Test plan

• uv run pytest — 18 passed. One pre-existing test
  (test_production_stdio_command_lists_tools) spawns the real server and
  requires a live Docker daemon; it fails identically on unmodified main on a
  daemon-less host (verified by stashing this change and re-running) and is
  unrelated to this diff.
• New tests/test_output_schemas.py (13 tests) exercises the real
  docker_to_dict — summary shape, full shape, the 500-char health-log cap,
  the image-ref fallback, overrides, networks — with faithful docker-py models,
  plus two end-to-end MCP calls proving list_containers defaults to the
  compact shape and run_container still returns the full object.
• uv run ruff check and uv run ruff format --check (the repo's devbox lint
  commands): clean.
• Live-daemon verification (read-only, over ssh:// to a 22-container
  host): list_containers default is 11,048 chars vs 102,965 full (89%
  smaller); 40 non-empty health-check outputs in full mode, 0 exceeding the
  cap; the format param is present on list_containers and absent from
  run_container.

Files

• src/mcp_server_docker/output_schemas.py — split docker_to_dict into
  summary/full builders; added format param and the health-log cap.
• src/mcp_server_docker/server.py — added format to the three list tools.
• tests/test_output_schemas.py — new direct tests of the real serializer.
• tests/test_server.py — updated the docker_to_dict mock to accept the new
  format kwarg.

[mecworks]

Proposed solution for #66: this PR adds the format parameter (default summary) and the 500-char health-log cap described there. Fixes #66.