fix(go-adk): truncate A2A session name on rune boundaries

[anxkhn]

What

extractSessionName in go/adk/pkg/a2a/executor.go derives a session name from the
first text part of an inbound A2A message and caps it at sessionNameMaxLength (20).
The cap used a byte-wise slice:

if len(tp.Text) > sessionNameMaxLength {
    return tp.Text[:sessionNameMaxLength] + "..."
}

Both len() and the [:20] slice operate on bytes, not runes. A first message longer
than 20 bytes that begins with multi-byte UTF-8 (CJK, emoji, accented Latin) gets cut
in the middle of a rune, producing an invalid-UTF-8 string.

This is only ~6-7 non-ASCII characters, so it is easy to hit for non-English users.
For example こんにちは世界 (7 CJK runes = 21 bytes) truncates to
こんにちは世\xe7\x95..., which is not valid UTF-8.

Why it matters

The derived name flows into state[StateKeySessionName] and is persisted via
CreateSession:

• On the Postgres session store, a text column rejects invalid UTF-8, so the
  session-create call fails and the whole turn is aborted with an error, a valid
  non-English first message can break the conversation.
• On SQLite, the invalid bytes are stored as mojibake and shown in the UI.

The fix

Count the cap in runes and slice on a rune boundary, so the result is always valid
UTF-8:

if utf8.RuneCountInString(tp.Text) > sessionNameMaxLength {
    return string([]rune(tp.Text)[:sessionNameMaxLength]) + "..."
}

Pure-ASCII behavior is unchanged (for ASCII, byte count == rune count), and the "..."
suffix is kept. The only additions are the unicode/utf8 import and a short doc comment
explaining why truncation is rune-aware.

Tests

Adds a table-driven TestExtractSessionName colocated in
go/adk/pkg/a2a/executor_test.go covering:

• nil message / no parts / short ASCII (unchanged behavior)
• ASCII exactly at the 20-rune limit (not truncated)
• ASCII over the limit (truncated with ...)
• multi-byte over the 20-byte limit but under the rune limit (must stay valid UTF-8)
• long multi-byte truncated on a rune boundary
• skipping an empty text part and using the first non-empty one

Each case asserts both the exact result and utf8.ValidString(got). The multi-byte
cases fail against the old byte-slice code (they return invalid UTF-8) and pass with the
fix; the ASCII cases pass either way.

Verified locally:

cd go && go test -race -run TestExtractSessionName -v ./adk/pkg/a2a/   # PASS (8/8)
cd go && go test -race -skip 'TestE2E.*' ./adk/...                     # ok, no regressions
cd go && go build ./adk/... && go vet ./adk/pkg/a2a/                   # clean

This is a small (<100 LoC) bug fix with unit coverage and no CRD/API change, so no E2E
change is needed.

[Copilot]

Pull request overview

Fixes UTF-8 corruption in A2A session name derivation by ensuring truncation happens on rune boundaries (preventing invalid UTF-8 that can break Postgres session creation), and adds focused unit tests to lock in correct behavior for ASCII and multibyte inputs.

Changes:

• Update extractSessionName to truncate based on rune count rather than byte count.
• Add a table-driven TestExtractSessionName validating exact outputs and utf8.ValidString(...) across ASCII + multibyte scenarios.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
go/adk/pkg/a2a/executor.go	Makes session-name truncation UTF-8 rune-aware to avoid invalid strings on multibyte input.
go/adk/pkg/a2a/executor_test.go	Adds unit coverage ensuring truncation correctness and UTF-8 validity for multibyte and ASCII cases.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.