fix: gate token-based compaction on latest prompt size, not event sum

[phillip-hicks-siro]

Fixes #473.

TokenBasedContextCompactor.shouldCompact summed usageMetadata.promptTokenCount across active events. Each of those values is the full prompt size of the call that produced it (system instruction + tools + entire history), so the sum re-counts the same history once per event and grows with call count rather than context size — and after a compaction, retained tail events still carry pre-compaction counts, leaving the sum permanently above any threshold (compact-every-call thrashing). Details and production impact in #473.

This ports the Python ADK's semantics (apps/compaction.py::_latest_prompt_token_count):

• use the most recently observed promptTokenCount — the measured size of the last real request
• fall back to a chars/4 estimate over the effective request contents (getContents, mirroring the content request processor) when no usage metadata is available

compact() and retention behavior are unchanged.

Testing: rewrote the threshold unit test against the new semantics and added: a latest-not-sum regression case, newest-metadata-wins, estimate fallback (over and under threshold), and the no-signal case. Full unit:core suite passes (954 tests).

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[phillip-hicks-siro]

@googlebot I signed it!

[Varun-S10]

Hi @kalenkevich, I have checked the issue and validated the proposed changes. I was able to reproduce the bug. Could you please review this PR?

[Varun-S10]

Hi @phillip-hicks-siro, We noticed that a few of the pre-checks are failing on this PR. Could you please resolve them?

[phillip-hicks-siro]

@Varun-S10 Pre-check error resolved!

[phillip-hicks-siro]

The windows-latest failure is unrelated to this change: tests/integration/a2a/basic/a2a_agent_test.ts failed at suite setup ("CLI exited prematurely with code 1" spawning the A2A server child process, test_case_utils.ts:332) before any test ran. The same suite passed on ubuntu and macos in this run, all 2307 individual tests passed on windows, and this PR touches only the token-based compactor and its tests. Could a maintainer re-run the windows job? Happy to rebase instead if preferred.