feat(agent): per-call timeouts, parallel-tool concurrency cap, max-iteration signal (Wave 5)

[urmzd]

Wave 5 — GA hardening (bounded slice)

All changes are in the agent package and stay green: go build ./..., go vet ./agent/, and go test ./agent/ pass (race-clean on the affected tests).

What landed

(a) Per-call timeouts

• AgentConfig.LLMTimeout / AgentConfig.ToolTimeout (time.Duration, 0 = none) + WithLLMTimeout / WithToolTimeout options.
• A child context.WithTimeout wraps the provider call inside getAssistantMessage and each tool step inside executeOneTool.
• A slow provider that honours the deadline surfaces a transient *types.ProviderError wrapping context.DeadlineExceeded (errors.Is(err, ErrProviderFailed) and IsTransient(err) both hold).
• A slow tool surfaces a deadline-exceeded tool error — including the case where the tool ignores ctx and completes after the deadline (checked via stepCtx.Err() after execution).

(b) Parallel-tool concurrency cap

• AgentConfig.MaxParallelTools (int, 0 = unlimited) + WithMaxParallelTools.
• executeToolsConcurrently bounds the fan-out goroutines with a buffered-channel counting semaphore.
• The durable-runner path (non-Noop StepRunner) still runs tools sequentially in the caller's goroutine — unchanged.

(c) Max-iteration truncation signal

• types.ErrMaxIterations was defined but never emitted. runLoop now tracks pendingWork: when it breaks on the iteration cap while the last assistant turn still had pending tool calls, it emits types.ErrorDelta{Error: ErrMaxIterations} so consumers can distinguish "truncated" from "finished".
• A clean natural finish (text-only turn or empty response) clears pendingWork and does NOT emit the error.

Tests (agent/limits_test.go, table-driven)

• LLM timeout fires, is transient, wraps DeadlineExceeded; disabled-by-default path stays clean.
• Tool timeout fires for a ctx-respecting tool and for a ctx-ignoring slow tool; fast tools unaffected.
• Concurrency cap: max-observed concurrency never exceeds N (probe tool with max-observed assertion); unlimited (0) runs all in parallel; cap=1 serializes and still returns every result.
• ErrMaxIterations emitted on a truncated run; NOT emitted on a clean finish nor when the assistant lands its final text turn exactly on the cap.

Deferred (TODO, beyond this slice)

• EventStream backpressure policy.
• Subagent depth limit + Logger/Metrics/StepRunner propagation into subagents.
• NewAgent -> (*Agent, error) (currently panics on invalid handoff config).
• Tracing coverage.
• Semver policy.

Stacked on #38 (foundation) — merge after #38.

[urmzd]

Superseded by #44, which was squash-merged into main (53a6aff) and contains every change from this branch — a dry merge of this branch into main is a no-op. main is green (build/vet/golangci-lint, 39 packages, and 8/8 live validation on gpt-4o-mini). Closing as redundant.