Docs truth: commands that do not run, snippets that do not compile

tracker: AGS-39

Two independent doc audits found systematic doc↔code drift. The pattern is not typos — whole chapters document APIs that were refactored away, and the primary run commands fail.

• Most documented mvn exec:java commands cannot run. exec-maven-plugin is not declared in the root pom at all, and Flink is provided scope while exec:java defaults to the runtime classpath — so every Flink-job example dies with NoClassDefFoundError. The repo already knows: TEST_REPORT.md records using -Dexec.classpathScope=test, and one example documents the workaround in its own javadoc.
• The doc presenting "the banking agent as Java" runs a generic Ollama chat bot with no tools whose own javadoc says "No Flink".
• flink run "$JAR" <class> in the market scripts passes the class as a program arg, so it launches the manifest Main-Class instead.
• docs/reference/external-state-storage.md is entirely fictional — every class and method it documents is absent repo-wide, and its DDL matches neither the implementation nor sql/schema.sql. Recommend deleting it.
• docs/reference/storage-architecture.md has the status table inverted: Postgres/Qdrant/pgvector are marked "Planned" while all three ship and are ServiceLoader-registered, and it documents a RedisShortTermStore that does not exist while never mentioning the real Flink-state-first memory.
• The storage-backends guide teaches an obsolete registration flow ("add a case to the factory switch") and never mentions the actual ServiceLoader + META-INF/services mechanism.
• Wrong env var names (POSTGRES_URL vs the real AGENTIC_FLINK_POSTGRES_URL), wrong Qdrant config keys and default port, and a ContextItem(String, String) constructor used 8 times that does not exist.
• TEST_REPORT.md is stale: claims 487/487 on Java 17; actual is 770 tests, 0 failures on JDK 21.

Clean and should not be rewritten: docs/examples/incident.md (zero findings) and docs/guides/creating-tools.md (~95% accurate).

Roadmap epic. Detailed findings, file:line citations and sub-tasks live in the AI-grind tracker (project AGS), which is the source of truth for this work. Local dev — build, test, tracking, profiling — is managed with AI-grind.

Tracked as AGS-39 in the devtools-mcp tracker.

Start with the AGS-39 tracker, which contains the detailed findings and file-line citations. Inspect the root pom, the named files under docs/reference, the market scripts, and TEST_REPORT.md, then compare each claim with the implementation and sql/schema.sql. Done means the documented commands, APIs, configuration names, storage registration, status tables, and test totals match the repository, while the fictional guide is removed and the clean guides remain unchanged.