feat: add unload on_agent_switch builtin hook

[dgageot]

Re-implementation of #2660 as a builtin hook instead of explicit runtime calls. Opened to compare the two designs side by side.

What

Adds an unload on_agent_switch builtin hook that asks every model on the previous agent to release its resources when the runtime hands control to another agent. Today only Docker Model Runner ships a provider.Unloader; everything else is silently skipped, so wiring the hook on a cross-provider chain is harmless.

Opt in via the standard hook YAML — no new flag, no auto-injection:

agents:
  coder:
    model: qwen3-large
    handoffs: [reviewer]
    hooks:
      on_agent_switch:
        - type: builtin
          command: unload
  reviewer:
    model: qwen3-coder
    handoffs: [coder]
    hooks:
      on_agent_switch:
        - type: builtin
          command: unload

DMR's _unload URL is derived from base_url (replacing the trailing /v1); a per-provider unload_api field overrides it for non-standard deployments. Errors are logged and swallowed, each call is bounded to 10 s — agent switching never blocks on a slow or unreachable engine.

How it compares to #2660

	#2660	this PR
Trigger sites	3 explicit r.unloadOnSwitch(...) calls in agent_delegation.go and runtime.go	1 hook entry, dispatched by the existing on_agent_switch machinery
Discoverability	private runtime method	shows up next to cache_response, redact_secrets, … in the schema and docs
Composability	hard-coded	users can interleave their own on_agent_switch hooks
Runtime LOC	~40	~25
Provider packages touched	provider, dmr, openai, base	provider, dmr

The single trade-off vs. #2660: LocalRuntime.SetCurrentAgent (the TUI agent picker) doesn't currently fire on_agent_switch hooks, so the hook-based version doesn't cover that path. Closing the gap is a one-line change in SetCurrentAgent that benefits every other on-switch hook too — happy to do it as a follow-up if we go with this approach.

Files

Feature

• pkg/model/provider/provider.go — Unloader interface
• pkg/model/provider/dmr/unload.go (+ tests) — DMR Unload, default URL derivation, override resolution, HTTP POST
• pkg/model/provider/defaults.go — plumb unload_api from ProviderConfig into the model's provider_opts
• pkg/config/latest/types.go (+ tests) — ProviderConfig.UnloadAPI field, ModelConfig.UnloadAPI() accessor
• pkg/runtime/unload.go (+ tests) — BuiltinUnload const + the closure registered on LocalRuntime
• pkg/runtime/runtime.go — registration in NewLocalRuntime

Schema, docs, example

• agent-schema.json — unload_api field, unload builtin in the command-list blurb
• docs/configuration/hooks/index.md — built-ins table row + on_agent_switch section
• docs/providers/dmr/index.md — "Unloading models on agent switch" section, single-tenant warning callout
• docs/providers/custom/index.md — unload_api row in the provider properties table
• examples/unload_on_switch.yaml (+ examples/README.md row)

Other things in this branch

The third commit (fix lint: …) cleans up 29 pre-existing lint offences that were sitting in the tree (3 govet/inline, 20 modernize/slicesbackward, 6 Lint/SlogContextual). I noticed them while running task lint against my changes and fixed them in a separate commit — happy to split into a separate PR if preferred.

One thing worth flagging: golangci-lint --fix rewrote truncateOldToolContent in pkg/session/session.go from for i := ...; i-- to for _, v := range slices.Backward(...) and changed msg := &result[i] to msg := &v. The latter points to the loop-local copy, so msg.Content = … was silently mutating the copy and the function no longer truncated anything. I corrected it to for i := range slices.Backward(result) { msg := &result[i] }. Worth being aware of when running --fix elsewhere.

Validation

• task build ✅
• task test ✅ (full suite)
• task lint ✅ — 0 offences (tree was at 29 before)
• New tests cover: builtin behaviour (happy path, error swallowing, same-agent no-op, registry lookup), DMR URL derivation/override, response body draining for connection reuse, unload_api plumbing through provider-config merge.

[rumpl]

There are a lot of unrelated changes in this PR and a lot of conflicts