fix(compression): preserve network/auth abort flags across cooldown re-entry (#29559)

[srojk34]

Summary

• PR fix(compression): abort + preserve context on transient network summary failure (#29559, #25585) #51881 fixed context loss on transient network failures by aborting compression instead of dropping the middle window ([Bug] Context compression silently drops summary on connection error, causing total context loss for long tasks #29559, bug(compression): failed summaries should not discard conversation context #25585).
• However, compress() unconditionally resets _last_summary_network_failure and _last_summary_auth_failure at the top of every call. When a second auto-compress fires during the 30-second failure cooldown, _generate_summary() returns None from the cooldown early-return without re-asserting these flags — so the abort guard sees False and falls through to the destructive static-fallback path, dropping the middle context window.
• Both flags are already cleared on success (_generate_summary lines 1697-1698), so the eager reset was redundant and harmful. This PR removes it so the flags persist during cooldown and the abort guard fires correctly on re-entry.

Repro scenario

1. Network blip causes ConnectionError during auto-compress → first compress() correctly aborts (PR fix(compression): abort + preserve context on transient network summary failure (#29559, #25585) #51881) ✅
2. User sends another message (adding tokens) → second compress() triggers within 30s cooldown
3. compress() resets _last_summary_network_failure = False → _generate_summary() hits cooldown guard, returns None without re-setting flag → abort guard sees False → middle window dropped ❌

Fix

Remove the eager _last_summary_auth_failure = False / _last_summary_network_failure = False reset at the top of compress(). The flags are already cleared on success, and persisting them during cooldown is safe and correct.

Test plan

• TestCooldownReentryAbort::test_network_failure_cooldown_reentry_still_aborts — ConnectionError → first compress aborts → second compress during cooldown also aborts (not fallback)
• TestCooldownReentryAbort::test_auth_failure_cooldown_reentry_still_aborts — same for 401 auth failures
• All 122 existing test_context_compressor.py tests pass (zero regression)

$ pytest tests/agent/test_context_compressor.py -q
122 passed in 7.97s

[teknium1]

Thanks for the focused fix. I verified the premise against current main and did not find blocking problems.

Current main still resets _last_summary_auth_failure / _last_summary_network_failure at agent/context_compressor.py:2402, while _generate_summary() returns during cooldown at agent/context_compressor.py:1476 without reasserting them. The abort guard at agent/context_compressor.py:2530 then misses the prior network/auth failure and falls through to the fallback path. I also reproduced that locally with .venv/bin/python: first ConnectionError compression preserved all 12 messages, second cooldown re-entry dropped to 6 messages with _last_summary_fallback_used=True.

The PR’s shape matches the bug class: it removes the eager resets and adds network plus auth regression tests for second-call cooldown re-entry. It also preserves the success path, where _generate_summary() already clears both flags at agent/context_compressor.py:1697.

No design-fit issues found: no new tools, env vars, schema surface, or cache-breaking prompt changes. Salvageability looks high; this should be mostly mechanical to cherry-pick.

Automated hermes-sweeper review.

[kshitijk4poor]

Landed on main in commit 82ac7e1 (authorship preserved) via batch salvage #56197. The eager flag-reset removal now correctly preserves the abort across cooldown re-entry — verified with a mutation-checked regression test. Thanks @srojk34!