fix(memory): skip non-text parts and use millis timestamps in memory search

[rootkiller6788]

Summary

InMemoryMemoryService.searchMemory had two bugs:

1. Crash on non-text parts. It called Part.text().get() unconditionally for every part in a stored event. When a session contains tool activity, the stored events include functionCall / functionResponse parts whose text() Optional is empty, so the next loadMemory search threw NoSuchElementException and the memory tool failed.

2. Wrong timestamp unit. formatTimestamp used Instant.ofEpochSecond, but Event.timestamp() is epoch milliseconds (Event.build() defaults to Instant.now().toEpochMilli()), so returned memory timestamps were ~1000× in the future.

Changes

• Skip non-text parts when building the keyword index (only text parts are searchable).
• Format memory timestamps from epoch millis via Instant.ofEpochMilli.
• Add InMemoryMemoryServiceTest covering both regressions.

Verification

• New tests pass: InMemoryMemoryServiceTest (2 tests).
• AgentWithMemoryTest still passes.
• Full core module suite: 1776 tests, only pre-existing Windows path-separator failure in LocalSkillSourceTest.testListResources (fails on untouched HEAD too).

[sherryfox]

Hi @rootkiller6788, thank you for your contribution and for taking the time to submit this PR.

You're right that InMemoryMemoryService.searchMemory has two bugs. The first fix unfortunately duplicates #1465, which should land soon. The second is definitely needed.

One ask for the scope then: once #1465 lands, could you rebase this down to just the formatTimestamp change? That drops the conflict and leaves a small, focused change that's easy to land — right now both PRs add the same new test file.

[rootkiller6788]

Thanks for the review, @sherryfox. I've narrowed the PR down to just the formatTimestamp fix as you suggested.

• Dropped the non-text part guard from this branch — it's covered by fix(memory): guard parts without text in searchMemory #1465, so keeping it here would duplicate the fix.
• Removed InMemoryMemoryServiceTest.java from this PR. Since fix(memory): guard parts without text in searchMemory #1465 also adds a file under that name, I moved the timestamp regression test into its own InMemoryMemoryServiceTimestampTest.java so there's no conflict on the test file once fix(memory): guard parts without text in searchMemory #1465 lands.
• The production change is now a single line: formatTimestamp reads the event timestamp as epoch milliseconds via Instant.ofEpochMilli, matching Event.timestamp().

The timestamp test passes against the core module (1 test, no failures). Could you please take another look?

[sherryfox]

Two mechanical things @rootkiller6788: check-commit-count is red at 2 commits, and the title still says "skip non-text parts" — worth retitling, since the title becomes the commit message and release-please turns it into the changelog entry.

[rootkiller6788]

Fix: memory search timestamps were parsed as seconds, not millis

InMemoryMemoryService.formatTimestamp() was decoding timestamps with Instant.ofEpochSecond(...). The timestamps going
through it are epoch millis, so the formatter read them as seconds and search result times came out wrong — anything
recent showed up as a date in 1970.

Changed it to Instant.ofEpochMilli(...), which is what the values actually are.

Tests

• Added InMemoryMemoryServiceTimestampTest covering the timestamp formatting path.

This is the narrowed version of the earlier fix — just the timestamp bug, nothing else.