feat: workflow classifier for definitions and execution search

[bradyyie]

Summary

Adds a workflow classifier — a tag that distinguishes kinds of workflow definitions and executions (plain workflows vs AgentSpan agents vs any future kind), derived from WorkflowDef.metadata:

1. explicit metadata.classifier (non-blank) wins;
2. otherwise agent when the AgentSpan stamp (agent_sdk / agentDef) is present;
3. otherwise workflow (plain, untagged def).

This is the OSS equivalent of the classifier support in Orkes Conductor, adapted to the OSS IndexDAO architecture. It pairs with conductor-agentspan:0.4.2 (agentspan-ai/agentspan#306), which uses it for classifier-based agent filtering.

Changes

common

• New WorkflowClassifier utility (+ unit tests).
• WorkflowSummary gains a classifier field (proto field 23), populated at index time from the execution's workflow definition; protos regenerated.

REST

• GET /api/metadata/workflow?classifier=agent|workflow|<tag> — filters definitions by derived classifier.
• GET /api/workflow/search and /search-v2 accept &classifier= (comma-separated), folded into the structured query as classifier IN (...) — the IndexDAO contract is unchanged, and raw query=classifier IN (agent) works too.

Index backends

• Postgres: Flyway V17 adds a classifier column + (classifier, start_time) index on workflow_index; written on index, whitelisted in PostgresIndexQueryBuilder. Filtering for the workflow token also matches pre-migration NULL rows.
• SQLite: same, via V5 + SqliteIndexDAO/SqliteIndexQueryBuilder.
• ES7 / ES8 / OS v2 / OS v3: classifier keyword mapping on workflow documents (required for ES8's dynamic: false); the query parsers pass field names through generically, so no parser changes.

AgentSpan embedding

• conductor-agentspan bumped 0.3.0 → 0.4.2 (first published version with classifier-based agent filtering: &classifier= on /api/agent/executions and /api/agent/executions/search).
• Fixed a packaging gap: nothing put :conductor-agentspan-server on the server classpath after the embedded glue was split out of the ai module, so the embedded agent endpoints 404'd on a stock build. The module is now included; activation remains gated on conductor.integrations.ai.enabled.

Testing

• Unit tests: WorkflowClassifierTest, classifier cases in PostgresIndexQueryBuilderTest / SqliteIndexQueryBuilderTest (incl. legacy-NULL-row semantics), REST controller tests for the new params.
• Integration: classifier index/search scenarios in PostgresIndexDAOTest (Testcontainers) and SqliteIndexDAOTest.
• Live end-to-end against a local Postgres-backed server with AgentSpan embedded: agent deployed/run via the Python SDK (agent_sdk stamp → classifier=agent), explicit metadata.classifier=pipeline def, untagged def; verified ?classifier= on /metadata/workflow, &classifier= and query=classifier IN (...) on /workflow/search, /api/agent/list, /api/agent/executions?classifier=agent, /api/agent/executions/search?classifier=, and that hand-inserted pre-migration NULL rows match the workflow token without leaking into agent.
• spotlessApply clean; full-repo main+test compilation passes.

Notes for reviewers

• Untagged executions index the literal workflow (never NULL going forward) so filtering is plain equality across every backend; the NULL-handling in the Postgres/SQLite builders exists only for rows indexed before the migration. Pre-existing ES/OS documents won't match classifier filters until re-indexed.
• search-by-tasks endpoints are unchanged — classifier is a workflow-level concept.