[Frankie-Xu]

Summary

Retrieval assembly now keeps raw content and returns structured evidence in composed. Update the connected-image regressions to match current main: an image is inserted at its matching placeholder, while an image without a reachable placeholder is skipped without opening storage.

The placeholder regression reads a valid tiny synthetic PNG through a mocked storage boundary and checks the exact image bytes and surrounding text order. This is a test-only change; current main is merged with normal history. Related to #206.

Verification

Validated against main f9d61119 using Python 3.11.15 and locked pytest 9.0.3:

• Original PR head: python -m pytest -p no:postgresql packages/shared-python/shared/tests/test_asset_inline.py -q — 9 passed.
• Original assertions after merging current main — 2 failed, 7 passed; both failures reproduced the obsolete flattened-content expectations.
• Updated branch: python -m pytest -p no:postgresql packages/shared-python/shared/tests/test_asset_inline.py packages/shared-python/shared/tests/test_evidence_compose.py -q — 20 passed.
• ruff check --no-cache packages/shared-python/shared/tests/test_asset_inline.py with Ruff 0.15.11 — passed.
• Pyright 1.1.408 with a temporary basic configuration scoped to this test file and the checkout's shared import path — 0 errors.
• Independent source review and git diff --check against current main — passed.

Dependencies came from a read-only locked API/shared environment plus isolated locked BeautifulSoup/soupsieve wheels. Tests used synthetic settings and fixtures. Complete API/worker suites and external parsing, model, or storage services were not exercised. GitHub workflow approval remains an external gate; empty check rollups do not establish a CI pass.

Deployment Notes

No production code, environment variables, migrations, queues, storage formats, or deployment changes. The assertions follow current main's existing composition contract.

Checklist

• Tests were added or updated when behavior changed
• Public docs, examples, or OpenAPI contracts were updated when needed (not applicable)
• Database migrations are idempotent and safe to deploy (not applicable)
• Logs, errors, and validation paths avoid leaking secrets or user data (synthetic fixtures only)
• The pull request description explains any breaking or user-visible change

[Frankie-Xu]

Rebased onto current main (d6f45c7f, includes #410).

Kept both assemble-level image regressions from this PR:

• connected image without a placeholder (Connected image chunks silently dropped in assemble_retrieval_results #206 silent-drop lock)
• image inserted at the placeholder

Resolved the conflict with #410's allowed_chunk_types filter tests. Please approve Actions workflows for this first-time fork PR so CI can run.

[Frankie-Xu]

Rebased onto current main (9489fa2c, includes #414/#415). test_asset_inline.py had no further conflicts after #410. Still waiting on first-time fork Actions workflow approval.

[EricNGOntos]

Thanks. We're not landing test-only locks unless they pin a regression we're actively chasing. Closing this one.