ai: linked SQL and verdicts in the DeploymentAnalyzer input

[babltiga]

Part of #1029 (linked deployment changes epic — read it first for the big picture). Needs #1031.

Goal

Show the deployment analyzer the database changes a release actually ships. The analysis input
gains the linked queries' SQL and their prior AI verdicts, rendered under a separate size cap and
placed in a dedicated section of the deployment prompt. Fail-safe behaviour is unchanged.

Today ai/api/DeploymentAnalyzer.java:34-37 defines DeploymentAnalysisInput(organizationId, aiConfigId, provider, environment, version, commitSha, artifactRef, metadataContext, language);
deploygov/internal/DeploymentAnalysisListener.java:70-74 builds it and caps the metadata slice
with MAX_METADATA_CONTEXT_CHARS = 16_000 (:45, renderMetadata :96-104).
ai/internal/DefaultDeploymentAnalyzer.java:26-41 frames a %s-formatted text block and hands it
to strategy.analyze(framed, DbType.CUSTOM, null, language, aiConfigId) (:45) — there is no
{{placeholder}} template and no SystemPromptRenderer involvement for deployments, so "a new
prompt placeholder" here means a new %s slot in that block. Lines :36-38 already tell the model
to treat "schema or data migrations" as elevated risk, blind.

Design constraints agreed on the epic:

• ai imports nothing new. The context is assembled in deploygov through core.api and
  passed as a string; ai never learns what a deployment link is. deploygov → ai.api stays the
  only arrow (docs/05-backend.md:4326-4334).
• A lookup failure degrades to a placeholder line — never a failed analysis, never an exception
  out of the listener beyond the existing AiAnalysisException | AiAnalysisParseException catch.
• Prior verdicts are context, not authority: the prompt labels them as earlier analyses of the
  individual statements; the deployment's own verdict is still the model's.

Steps

1. ai/api/DeploymentAnalyzer.java — add String linkedChangesContext to
   DeploymentAnalysisInput (null or blank = none). Record, so the change is a compile-time
   fan-out: update DeploymentAnalysisListener, DefaultDeploymentAnalyzerTest,
   DeploymentAnalysisListenerTest, and any other constructor call (grep -rn "DeploymentAnalysisInput(").
2. deploygov/internal/DeploymentAnalysisListener.java — static final int MAX_LINKED_CHANGES_CONTEXT_CHARS = 24_000 beside the metadata cap, and a renderLinkedChanges
   that, for each link in sequence_order, reads QueryRequestLookupService.findById (SQL,
   queryType, status, datasourceId) and AiAnalysisLookupService.findByQueryRequestId
   (riskLevel, riskScore, summary, failed) and emits one block per link:
   #<seq> [<queryType>] status=<status> datasource=<name> prior_ai_risk=<LEVEL>/<score> on one
   line, the summary on the next, then the SQL — each SQL individually capped so one 50 kB
   migration cannot starve the rest, and the whole string truncated with the same
   "\n… (truncated)" marker renderMetadata uses. A missing analysis renders
   prior_ai_risk=(none), a failed one (analysis failed), a missing query (query no longer exists). Wrap the whole render in a RuntimeException catch that returns
   "(linked changes could not be loaded)" and logs at WARN. Both caps are constants (the
   deploygov: DeploymentReviewEscalationJob, reviewer nudges and DeploygovProperties #985 DeploygovProperties can absorb them later if anyone asks).
3. ai/internal/DefaultDeploymentAnalyzer.java — one new %s section between the release
   metadata and the risk guidance:
   Linked database changes (each is a governed AccessFlow query request; prior_ai_risk is an earlier per-statement analysis, not a verdict on this deployment): followed by
   blankToNone(input.linkedChangesContext()). Extend the guidance line to say that a linked
   change already rated HIGH/CRITICAL, a DDL against a production-like environment, or a deploy
   that ships no linked change but whose metadata mentions migrations, is elevated risk. Keep the
   null schema-context argument and its comment (:42-44) — the linked SQL is already inside
   framed; passing it twice bills the org twice.
4. docs/05-backend.md (:4326-4334) and docs/18-deployment-governance.md §3 (:153)
   — describe the new input field, both caps, the prompt section and the degradation rules;
   deploygov: linked-changes documentation sweep #1036 does the wider sweep, but the analyzer contract is documented in this PR.

Acceptance criteria

• DefaultDeploymentAnalyzerTest: the framed prompt contains the linked section verbatim when
  the context is present and (none) when null/blank; the strategy still receives null schema
  context; the rate limiter is still enforced first.
• DeploymentAnalysisListenerTest: rendering order follows sequence_order; per-SQL and total
  caps truncate with the marker; missing analysis / failed analysis / deleted query render their
  placeholders; a RuntimeException from a lookup yields the "could not be loaded" line and the
  analysis still completes; a request with no links passes null; the existing skipped / failed
  event paths are byte-for-byte unchanged.
• mvn -f backend/pom.xml -q test -Dtest='ApplicationModulesTest,ApiPackageDependencyTest'
  passes — ai.api has no new import, and ai has no deploygov dependency.
• Coverage ≥ 90 % lines / ≥ 80 % branches on the touched classes.