feat(truth): cutover to MAF runtime

[Cataldir]

Problem statement

Migrate the truth-* bounded context (4 services: ingestion, enrichment, HITL, export) from the legacy BaseRetailAgent runtime internals to the MAF-backed runtime introduced in #981. The truth layer has the largest adapter surface in this cutover (4 services with shared enrichment / export semantics), so it lands as the second context to validate the seam under realistic adapter load before the wider agent fleet migrates.

This is a hard sunset: legacy runtime imports are deleted in the same PR. There is no coexistence flag.

Required change

For each of truth-ingestion, truth-enrichment, truth-hitl, truth-export:

• Replace usage of lib/src/holiday_peak_lib/agents/ internals with the public surface backed by the MAF runtime from feat(lib): MAF runtime adapter behind the existing BaseRetailAgent surface #981.
• Update each main.py and the agent definition to use the MAF-backed seam — no MAF idioms leak into service code unless intentional.
• MCP tool registrations via FastAPIMCPServer remain unchanged in signature; only the underlying runtime changes.
• REST endpoints remain unchanged in shape; migration is internal.
• Per-service Foundry config (FOUNDRY_AGENT_ID_FAST, MODEL_DEPLOYMENT_NAME_FAST, FOUNDRY_AGENT_ID_RICH, MODEL_DEPLOYMENT_NAME_RICH, FOUNDRY_STREAM, PROJECT_ENDPOINT) wired via Helm values + Key Vault refs.
• Three-tier memory bound per ADR-032 (PR docs(adr): ADR-029..033 - architecture contracts (canary, MCP A2A, OTEL, memory, UI on SWA) #978).
• MCP-only A2A enforced (ADR-030, PR docs(adr): ADR-029..033 - architecture contracts (canary, MCP A2A, OTEL, memory, UI on SWA) #978); hop counter wired.
• Delete pre-MAF imports across all four services in the same PR.

Affected components

• apps/truth-ingestion
• apps/truth-enrichment
• apps/truth-hitl
• apps/truth-export

Acceptance criteria (per spec — Acceptance gate per bounded context)

• Contract tests in lib/tests pass.
• All apps/truth-* unit and integration tests pass.
• Continuous eval baseline (per ADR-028 final schema, PR Add continuous agent evaluation engine #974) within maxDeltaPercent for each agent in this context (advisory until PR Add continuous agent evaluation engine #974 merges).
• AGC canary 5 → 25 → 50 → 100 succeeds in dev, then in staging (per ADR-029, PR docs(adr): ADR-029..033 - architecture contracts (canary, MCP A2A, OTEL, memory, UI on SWA) #978).
• No A2A HTTP detected by scripts/ci/lint_no_a2a_http.py (per ADR-030, PR docs(adr): ADR-029..033 - architecture contracts (canary, MCP A2A, OTEL, memory, UI on SWA) #978).
• Span attribute contract validated by App Insights workbook query (per ADR-031, PR docs(adr): ADR-029..033 - architecture contracts (canary, MCP A2A, OTEL, memory, UI on SWA) #978).
• No pre-MAF imports remain in apps/truth-ingestion, apps/truth-enrichment, apps/truth-hitl, apps/truth-export.
• Per-service Foundry config wired via Helm values + Key Vault refs for all four services.
• 24-hour steady-state hold satisfied before the next context's PR opens.

Risks and dependencies

Risk	Mitigation
Truth context has the largest adapter surface in the cutover — 4 services sharing enrichment/export semantics.	Land contract tests first under the seam introduced by #981; validate cross-service behavior with tests/e2e before opening the canary.
Eval baselines drift during cutover.	Baselines refresh nightly (per PR #974). R1 introduces no new behavior; baseline drift indicates a bug, not a feature.
MCP A2A introduces extra latency vs. previous A2A pattern.	Measured pre/post per service; tolerance P95 within 10 %. Outside tolerance halts the canary.
Reach-throughs into BaseRetailAgent internals from any of the 4 services.	Pre-flight grep across the four apps/truth-*/src trees for direct use of internal modules; blocks cutover until refactored.
Foundry agent ID provisioning fails for any of the 4 services.	Provisioning runs ahead of this PR; failure halts this context.

Blocked by:

• feat(lib): MAF runtime adapter behind the existing BaseRetailAgent surface #981 (lib MAF runtime seam).
• feat(crud-service): cutover to MAF runtime #982 (crud-service cutover — preceding context per the R1 cutover order).
• PR chore(lint): clear pylint baseline in truth-* services #979 (truth-* pylint baseline cleanup) — must merge before this PR opens, otherwise pylint clean-up risks blending into the MAF cutover diff.

Blocks: the next context (search-enrichment-agent) per the R1 sequencing rule.

Evidence links

• apps/truth-ingestion
• apps/truth-enrichment
• apps/truth-hitl
• apps/truth-export
• apps/truth-ingestion/src
• apps/truth-enrichment/src
• apps/truth-hitl/src
• apps/truth-export/src
• lib/src/holiday_peak_lib/agents/base_agent.py

ADR impact

• ADR-007 (memory tiers) — refined by ADR-032 (PR docs(adr): ADR-029..033 - architecture contracts (canary, MCP A2A, OTEL, memory, UI on SWA) #978).
• ADR-020 (product truth layer) — preserved; surface unchanged.
• ADR-024 → ADR-030 (PR docs(adr): ADR-029..033 - architecture contracts (canary, MCP A2A, OTEL, memory, UI on SWA) #978): MCP-only A2A enforced.
• ADR-029 (PR docs(adr): ADR-029..033 - architecture contracts (canary, MCP A2A, OTEL, memory, UI on SWA) #978): AGC canary gates the cutover.
• ADR-031 (PR docs(adr): ADR-029..033 - architecture contracts (canary, MCP A2A, OTEL, memory, UI on SWA) #978): span attribute contract validated.

Branch convention

feature/<issue-id>-truth-maf-cutover (per ADR-018).

BPMN process

%%{init: {'theme':'base', 'themeVariables': {
  'primaryColor':'#FFB3BA',
  'primaryTextColor':'#000',
  'primaryBorderColor':'#FF8B94',
  'lineColor':'#BAE1FF',
  'secondaryColor':'#BAE1FF',
  'tertiaryColor':'#FFFFFF'
}}}%%
flowchart LR
  A[Analyze Current Code] --> B[Design Change]
  B --> C[Implement on Issue Branch]
  C --> D[Open PR]
  D --> E[Validation and Fixes]
  E --> F[Merge to Main]
  F --> G[Monitor Workflows]
  G --> H[Close Issue and Cleanup]

Loading