refactor(agent): remove dead _budget_grace_call / _budget_exhausted_injected flags

[YiRaaaan]

Summary

Removes two unreachable flags — _budget_grace_call and _budget_exhausted_injected — along with their leftover consumer logic. Both are dead code: nothing in the tree ever sets either to True, so the branches that read them are unreachable and behaviorally equivalent to being deleted.

Why these are dead

The flags backed the one-turn budget grace mechanism introduced in #7983.

#8935 found that mechanism was broken: the flag was set after the while loop had already exited (so the loop could never re-enter to use the "grace" call), and worse, it blocked the _handle_max_iterations fallback from running — leaving final_response = None and surfacing empty responses on budget exhaustion. #8935 fixed this by calling _handle_max_iterations directly from the post-loop tail (now agent/turn_finalizer.finalize_turn), which already injects a summary request and makes one extra tool-less API call.

However, #8935 only removed the producer side (the setter + grace-message injection). The consumer side survived and was later carried into its current location by the god-file extraction refactors (extract run_conversation → agent/conversation_loop.py, extract __init__ → agent/agent_init.py):

• agent/conversation_loop.py — the or agent._budget_grace_call disjunct in the while condition, and the if agent._budget_grace_call: ... = False branch
• agent/agent_init.py — the _budget_grace_call and _budget_exhausted_injected initializers (the latter's setter was also removed in fix: budget-exhausted conversations now get a summary instead of empty response #8935)

Since neither flag is ever set True, ... or agent._budget_grace_call reduces to ... or False and if agent._budget_grace_call: is never taken — the loop always falls through to the normal iteration_budget.consume() path.

Changes

• agent/conversation_loop.py: drop the dead disjunct from the loop condition; collapse the if/elif so budget consumption is the single check.
• agent/agent_init.py: remove both flag initializers and their (now-stale) comment block.
• tests/run_agent/test_run_agent.py: remove the obsolete TestBudgetPressure::test_grace_call_flags_initialized (asserted the removed flags init to False).
• AGENTS.md: drop the stale or self._budget_grace_call line and "one-turn grace call" wording from the loop pseudocode.

Net: 4 insertions, 28 deletions. No behavior change.

Verification

• python -m py_compile on all touched modules — clean.
• grep -rn "_budget_grace_call\|_budget_exhausted_injected" across the tree — zero remaining references.
• pytest tests/run_agent/test_run_agent.py -k "budget or iteration or grace or max_iter or loop" — 13 passed; full file collects cleanly (371 tests).

🤖 Generated with Claude Code

[YiRaaaan]

Closing as a duplicate. #32597 already removes both _budget_grace_call and _budget_exhausted_injected (among other unused code), and #38282 takes the opposite approach — reviving the grace-call path — and is already tagged needs-decision. Whether to remove or revive this code is best decided on those existing PRs, so this one is redundant. Apologies for the noise, and thanks!