[mutianf]

Follow-up to #14681, from reviewing it after it merged. Three independent fixes.

1. Two drifting resolution rules (the main one)

DivertingUnaryCallable.resolveSessionLoad (per RPC) and ClientConfigurationManager.divertsAnyTrafficToSessions (client-wide) answered "does traffic go to sessions?" with separately maintained rules that had already diverged:

	per-RPC	client-wide
diversion config present	authoritative, flat load ignored	ORs with flat session_load
per-scope map	only this method's key	any key, incl. methods the client never serves

So per_scope_diversion { "method:Bigtable.MutateRows": {session_load: 1.0} } made areSessionsRequired() true — flipping the sessions_required / sessions_compatible feature flags sent to the server and keeping the session config and channel pool alive — while DivertingUnaryCallable diverted exactly zero requests, because no callable asks for that scope. Same for session_load: 1.0 plus a CheckAndMutateRow-only map.

Both now go through a new SessionScope enum, which owns the four scopes this client serves, whether each honours the deprecated flat session_load, and the single resolution rule (moved verbatim out of DivertingUnaryCallable). ShimImpl passes SessionScope.READ_ROW instead of a String + boolean pair.

2. The default config shadowed the flat session_load

bigtable-default-client-config.textproto carried a zero-valued per_scope_diversion for ReadRow and MutateRow. Since a diversion configuration is authoritative once present, those entries forced every scope to 0 in any config derived from the default:

• the initial normalizeConfig(defaultConfig) before the first GetClientConfiguration response,
• the handleFailedFetch → setClientConfiguration(defaultConfig) path when the validity window expires during a config-service outage,
• anything merged from the bigtable.internal.client-config-override sys-prop, including the bigtable-session-it Maven profile.

This surfaced as 10 ClientConfigurationManagerTest failures the moment fix 1 made the two checks agree — they had been passing only because the client-wide check disagreed and reported sessions as on. The entries were purely documentary (an absent key and an absent diversion config both resolve to 0), so they are removed.

3. ReadModifyWriteRow could return null on the session path

The extractor was r -> r.hasRow() ? r.getRow() : null, so a SessionReadModifyWriteRowResponse without a row handed user code a null the classic path never produces (EnhancedBigtableStub always builds a Row from the possibly-unset resp.getRow()). Unlike ReadRow — where a missing key legitimately means null — ReadModifyWriteRow always mutates and always yields a row, so it now extracts unconditionally. Both extractors are named constants with tests pinning the split.

Also forwards the per-method retry settings to the default call context for CheckAndMutateRow and ReadModifyWriteRow, matching the existing ReadRow and MutateRow wiring.

Testing

New SessionScopeTest (7 cases) covers the resolution rule and the agreement invariant; new cases in ShimImplTest pin the ReadRow-null / ReadModifyWriteRow-non-null split, and ClientConfigurationManagerTest gains one asserting the bundled default never shadows a flat session_load.

DivertingUnaryCallableTest, ClientConfigurationManagerTest, ShimImplTest, SessionDeadlineTest, SessionScopeTest, RowBuilderShimTest, ReadModifyWriteRowShimTest — 57 tests, all passing.

[gemini-code-assist]

Code Review

This pull request introduces a new SessionScope enum to encapsulate per-method diversion scopes and their session load resolution rules. This refactoring unifies the diversion logic previously split across DivertingUnaryCallable and ClientConfigurationManager, ensuring consistency between per-RPC routing and client-wide session management. Additionally, retry settings are now correctly propagated to the default call context for check-and-mutate and read-modify-write row callables, and corresponding unit tests have been added or updated. There are no review comments, so I have no feedback to provide.