Restore loop guards: repeat-call/output breakers, recoverable-failure headroom, terminal-inference halt, autonomous iter-cap lift

[senamakel]

Part of the v0.58.7 → HEAD TinyAgents-migration audit (see parent issue). Restores loop-guard behaviors deleted with tool_loop.rs that have no equivalent in the tinyagents crate (verified against crate source — RepeatedToolFailureMiddleware tracks failures only).

Dropped behaviors

1. RepeatCallGuard ([Harness] Add circuit breaker for repeated identical tool calls #4088): halt after 3× identical successful tool-call batches — a model looping on a no-op tool now burns the full iteration budget.
2. RepeatOutputGuard ([Harness] Agents enter unproductive loops (repeated tool calls, no progress) #4095): halt after 4× identical tool outputs.
3. Polling-tool exemption: tools legitimately called repeatedly (status polls) were exempt from both guards.
4. Recoverable-failure headroom collapsed: legacy gave transient/timeout failures extended headroom (8 identical / 12 varied) before halting; the new RepeatedToolFailureMiddleware thresholds are 3/6 (src/openhuman/tinyagents/middleware.rs:1680,1815) — long-running retry-appropriate operations now halt early.
5. Terminal delegated-inference fast-halt (Cascade error on agentic task — Plan, Run Code, and Tools Agent all failing in sequence #3104): budget-exhausted / config-rejection delegation failures halted the loop on first occurrence; gone — only coarse triage-level budget classification remains.
6. Policy-denied fast-trip weakened 2 → 3 repeats.
7. Autonomous iteration-cap lift is a dead knob: autonomous_iter_cap() has zero readers, while task_dispatcher/executor.rs:199 and skill_runtime/run_machinery.rs:228 still call with_autonomous_iter_cap expecting the lift (legacy ops/loop_.rs:73-77). Autonomous task/skill subagents now stop at the normal per-agent cap (e.g. 10) instead of TASK_RUN_MAX_ITERATIONS.

Note: the ~10 legacy tests covering 1–3 and ~10 covering 5 were deleted with the code (see test-parity issue).

Fix plan

1. Re-implement 1–3 as a seam middleware (e.g. RepeatProgressMiddleware in src/openhuman/tinyagents/middleware.rs): track (tool, args-hash) for successes and output-hash across rounds; on threshold, inject the legacy halt instruction / pause via SteeringCommand::Pause with the standard checkpoint text. Port the polling-tool exemption list.
2. Restore differentiated thresholds in RepeatedToolFailureMiddleware: recognize recoverable classes (timeout/transient markers, reusing the feat(agent): classify tool-call failures + surface on web event — rework of #4413 (backend) #4445 classifier) and give them the legacy 8/12 headroom; keep 3/6 for other failures; restore the 2-repeat fast-trip for POLICY_DENIED_MARKER.
3. Re-add the terminal delegated-inference classifier as a first-halt check on delegation-envelope failures.
4. Wire autonomous_iter_cap() back in where effective_max_iterations is computed for subagent runs (subagent_runner/ops/graph.rs / tinyagents/mod.rs:183) — or delete the setters and thread an explicit cap parameter (preferred: explicit parameter; the task-local was already fragile, and the detached-spawn issue shows why).
5. Restore the deleted tests for each guard as middleware unit tests + one loop-level e2e per guard.

Acceptance criteria

• Identical successful call 3× / identical output 4× halts with a resumable checkpoint; polling tools exempt.
• Transient failures get extended headroom; policy-denied trips at 2.
• Task-dispatcher and skill runs execute up to TASK_RUN_MAX_ITERATIONS again.