feat(agent): add registered business outcome mapping and outcome consistency fixes (T38)

[konlue]

Describe what this PR does / why we need it

Second and closing part of T38 (#141), following the "keep it simple" agreement: a registered business-outcome mapper is now the only component allowed to reinterpret a successful raw tool result as a business failure (no global string guessing), the reserved ToolOutcome.BUSINESS_FAILED vocabulary gets its first producer, and three outcome-consistency gaps found along the way are closed.

Does this pull request fix one issue?

Fixes #141

Describe how you did it

• BusinessOutcomeMapper (functional interface) + BusinessOutcomeMapperRegistry mirroring ToolEffectRegistry: builder().declare(toolName, mapper), fromTools(...) picking up BusinessOutcomeAware tools, merge(). The default registry is empty and results of unregistered tools are never inspected (a test feeds a {"error": true} payload to an unregistered tool and asserts it stays a plain success).
• The registry is applied at the single point between tool execution and ToolCallResponse.from(...): when a registered mapper marks a successful raw result as error, the response is built with status ToolOutcome.BUSINESS_FAILED — the vocabulary reserved in PR feat(agent): unify tool failure status contract and remove null-status error paths (T38) #159 now has its first producer.
• isError() recognizes BUSINESS_FAILED as a failure, so the model-visible message derived from the response carries tool.error=true — conversions no longer lose the outcome (one of the issue's acceptance criteria).
• LargeResultEvictionInterceptor keeps the original response metadata instead of dropping it when evicting a large result.
• ToolEmulatorInterceptor returns success(...) instead of the legacy of(...): an emulated result is a successful outcome and must not look like a failure to retry interceptors.
• ToolRetryInterceptor's exhausted-retries exit carries forward the last failure's structured failureKind/exceptionType entries.

Describe how to verify it

• New tests: BusinessOutcomeMapperRegistryTest (6 cases: no-guessing, declare, fromTools, merge, null-mapper rejection, argument validation), AgentToolNodeBusinessOutcomeTest (4 cases run end to end through the node, asserting BUSINESS_FAILED + metadata + model-visible tool.error), LargeResultEvictionInterceptorTest (metadata preservation); plus one new retry case and one emulator status assertion.

• Run:

  ./mvnw -pl argi-agent-framework -am test -DskipTests=false \
    -Dtest='BusinessOutcomeMapperRegistryTest,AgentToolNodeBusinessOutcomeTest,ToolRetryInterceptorUnitTest,ToolEmulatorInterceptorTests,LargeResultEvictionInterceptorTest' \
    -Dsurefire.failIfNoSpecifiedTests=false

• Full regression: graph-core 701 + agent-framework 1122 tests; only the known pre-existing Windows-environment failures remain (the FileSystem* symlink / virtual-root family and the skills/prompt-template cases reproduce on clean main, as documented in feat(agent): unify tool failure status contract and remove null-status error paths (T38) #159).

Special notes for reviews

• Behavior changes: emulated results now carry SUCCESS_STATUS (previously no status, so retry interceptors retried them as failures), and isError() returns true for BUSINESS_FAILED.
• BUSINESS_FAILED responses are not SUCCESS_STATUS, so retry interceptors retry them like other failures; if a business failure should be terminal for the retry policy, the mapper can keep returning a success result. Happy to adjust.
• "evicted_to_filesystem" intentionally stays outside the ToolOutcome vocabulary (it is a success variant, not a failure); let me know if you would rather fold it in or map it to SUCCESS.
• MCP adapter error mapping into this protocol is untouched (the adapter modules are new since the refactor) and can follow up separately.