test: union and dedup prior refs in merge_staged_sections (PRPUNDIT-14) - #1329

Follow-up for the union-half pin: before = [prior, staged_ref_a], after = [staged_ref_a, staged_ref_b], expect [prior, a, b]. Confirmed locally that mutating if before or after: → if not before and after: at values.py:1054 now fails this test (drops staged_ref_b).

Heads-up — I updated this branch from main so it would pick up the skipped-status fix from #1357, and that surfaced a break. skipped-status is green now, but ruff went red:

F821 Undefined name `FrameworkAgentKB`
 --> src/hyperloom/inference_optimizer/tests/test_remote_recipe_v2.py:2020

FrameworkAgentKB was removed somewhere in the 207 commits this branch was behind — it doesn't exist anywhere on main any more. It's now PatchKB in orchestrator/knowledge/agent_kb.py, and the pre-existing test right above yours (line 2002) already uses it:

refs = PatchKB(sections).stage_patches(...)

So it should be a one-word change on line 2020:

staged_refs = PatchKB(sections).stage_patches([patch_a, patch_b], stack_index=2)

PatchKB is already imported in the file post-merge, so nothing else should need touching. I confirmed it's a real runtime break, not just lint — locally the test errors with NameError: name 'FrameworkAgentKB' is not defined (1 failed, 125 passed).

I ran the other four updated branches too and they're all clean (#1324 51 passed, #1327 17, #1328 74, #1330 44) — this is the only one the merge caught out.

Happy to push the rename myself if you'd rather (maintainerCanModify is on) — just say the word, otherwise it's yours.

Pushed the fix to this branch (acf74f665) — it turned out to be more than the one-word rename I described above, so here's exactly what changed and how I checked it.

Two renames landed in the 207 commits this branch was behind, not one:

So after fixing the class name the test still failed, first with NameError, then AttributeError: 'NoneType' object has no attribute 'knowledge' (because sections.staged("framework") no longer resolves), then KeyError: 'framework'. The diff is 8 lines, all inside your test — the class name, the two prior_* ref prefixes, the staged(...) lookup, the value key, and the three assertions.

merge_staged_sections itself was also refactored on main, which matters for what the test is pinning. It used to be:

combined = {...before...}
if before or after:
    combined[ref_key] = list(dict.fromkeys([*before, *after]))

and is now seeded from the staged side instead:

combined = dict(staged.knowledge)
for ref_key, before in current.items():
    after = combined.get(ref_key, [])
    ...
    combined[ref_key] = list(dict.fromkeys(str(ref) for ref in [*before, *after] if str(ref)))

I re-ran the mutation set against the new shape to make sure the test still catches a regression rather than just passing:

• after = [] (drop the newly staged side — the silent-data-loss direction) — red
• current = {} (skip the union loop, drop prior refs) — red
• dedup removed — red
• [*before, *after] → [*after] — red
• operands reversed — red

126 passed, ruff check and format both clean.

One thing I deliberately left alone: the comment above after_patches still describes the old if not before and after: shape, which no longer exists. The test's behaviour is right, but that comment now points at code that isn't there. Wanted you to word that one rather than guess at it.

The other four branches I updated are unaffected — #1324, #1327, #1328 are all green and #1330 is nearly there.