feat(runtimes): acp runtime launches a named command with no model

[lex00]

Every runtime Fountain has is an LLM coding agent. A deterministic program that wants a warm machine, a vault, a thread, a schedule and a human who can read what happened could not be launched: the runtime registry is a closed map, the model parser expects a provider/model_id, and a turn resolves inference credentials it will never use. This PR adds an acp runtime that launches a named command and speaks ACP to it over stdio like the four existing runtimes.

• Add Fountain.Runtimes.Command, a Managoat.Runtimes implementation for a command the agent names, and a Fountain.Runtimes facade that answers acp itself and delegates everything else to the library registry. Every runtime lookup, including skills installation, goes through the facade.
• Add agents.runtime_command, required for acp and refused on any other runtime, and make agents.model optional for acp only. Both refusals are 422s naming the field. The command is a free string, run as bash -lc inside the sandbox.
• Resolve no inference credential and pin no model on an acp turn, so model.failed cannot come from the pin path and a turn on an account with no credentials succeeds. turn.usage is null; credits price the turn by sandbox time only.
• Expose the field in the API, the console form (which disables the model input on acp) and the account export. Model suggestions return nothing for acp.
• Add a conformance test that runs a checked-in fixture ACP agent, a small Elixir program, as a real OS process through a whole turn, plus scripted-agent tests for tool-call blocks, the stop reason, interrupt and all three permission verdicts.
• Document the runtime in the catalog, docs/sdk.md, docs/concepts/permissions.md and the integrations pages, including what "the same isolation" means on a self-hosted runner.

Closes #1634

Stacked on #1637; merge that first.

Validation:

• Full suite on the four-branch stack: 4579 tests, 0 failures (core and ee), 144 (fountain_buzz), 33 (fountain_support).
• mix compile --warnings-as-errors, mix format --check-formatted, mix credo --strict, dialyzer, sobelow and deps.unlock --unused are clean. The prod release assembles.
• The SDK contract check, the TypeScript verifier and tests, the Python verifier and the conformance lint pass. scripts/docs-style.py is clean.
• Not run here: vale, okf and destink (not installed on the build machine).

[jhgaylor]

Split into a stack of eight, per the rule against big PRs. This branch's 1,760 lines are now #1832 → #1839, each based on the one before it, off main rather than on the labels branch:

	PR	Lines	What it is
1	#1832	+279	Fountain.CommandRuntime and the host dispatch for acp
2	#1833	+399	agents.runtime_command, and acp becomes a runtime you can name
3	#1834	+73	skills mount on a runtime the library does not know
4	#1835	+332	a turn on the acp runtime, end to end
5	#1836	+72	refuse a turn whose agent no longer carries a command
6	#1837	+151	the acp runtime in the agent form
7	#1838	+399	a real ACP program through a whole turn
8	#1839	+139	docs, CHANGELOG and SDK 1.25.0

Four things had to change rather than be copied, because this branch is stale against main:

• The model pin. main now has TurnMachine.acp_model/2 and CodexChatGPT.peer_auth/2; this branch still called RuntimeDispatch.acp_model/2 from the conversation server. The stack keeps main's call sites and makes TurnMachine.acp_model/2 delegate through the dispatch, so the acp → nil rule lives in one place.
• TurnMachine.open/4's spec dropped main's {:error, term()} clause (execution allowances). The stack keeps all four returns.
• The migration timestamp 20260906120000 sat behind five migrations already on main. Renamed 20260910120000.
• The console default model anthropic/claude-sonnet-4-6 is no longer a catalog entry and failed an existing test on main. Now claude-sonnet-5.

And two defects that only showed up once the pieces were run separately:

• conversation_server.ex went over its size pin. The file is line-pinned and only shrinks (Tracker: ConversationServer by subtraction #1369); the added comment put it at 2775 against a pin of 2774. Folded into the existing comment instead of raising the pin.
• The conformance test is flaky as written. "the turn completes on the agent's stop reason" waits on the turn row reaching completed, then asserts the conversation row is idle — which the server writes afterwards. It reproduced on the first seed once the timing shifted. test(runtimes): conformance, a real ACP program through a whole turn (#1634) #1838 waits on the row the assertion reads; 25 seeds clean.

Full suite on the stack tip: 4805 core tests, 144 fountain_buzz, 33 fountain_support, 0 failures. Leaving this open for you to close.

[lex00]

Superseded by #1832 to #1839; the diff over apps/ matches apart from the deviations listed above.

[jhgaylor]

All eight landed on main, plus a replacement for one of them:

	PR
1	#1832	Fountain.CommandRuntime and the host dispatch for acp
2	#1833	agents.runtime_command, acp becomes nameable, and the Swift SDK widening
3	#1834	skills mount on a runtime the library does not know
4	#1835	a turn on the acp runtime, end to end
5	#1836	refuse a turn whose agent no longer carries a command
6	#1863	the console form (replaces #1837, see below)
7	#1838	a real ACP program through a whole turn
8	#1839	docs, ADR 0049 and SDK 1.26.0

git diff between this branch's rebased tree and main over the five core modules is empty, so what shipped is the code this PR proposed.

Three things the split turned up that this branch would have carried in:

• The Swift SDK broke on a null model. FountainKit.Agent declared model: String on a Decodable struct, so an acp agent's explicit null threw valueNotFound — and because a page decodes whole, one such agent broke agents.list() for every Swift caller. Widened in feat(agents): agents.runtime_command, and acp becomes a runtime you can name (#1634) #1833 with tests that fail on the original declaration. No gate caught it: the contract checker does not compare a hand-written model's optionality against a nullable wire field.
• conversation_server.ex went over its size pin. The file only shrinks (Tracker: ConversationServer by subtraction #1369) and the added comment took it to 2775 against a pin of 2774.
• The conformance test was flaky as written, waiting on the turn row and asserting on the conversation row.

Two unrelated flakes were filed along the way: #1861 (Swift conformance timeout scenario) and #1862 (TeamTest co-tenant test dying on a shutting-down Horde supervisor).

Closing this in favour of the stack.