manual_slop/docs/reports/TIER1_REVIEW_phase9_dilemma_20260620.md

Status Report: result_migration_baseline_cleanup_20260620 — Phase 9 Dilemma

Date: 2026-06-20 Track: result_migration_baseline_cleanup_20260620 (Sub-Track 5 of 5 in the result_migration_20260616 umbrella) Author: Tier 2 (autonomous sandboxed run) Status: 9 of 14 phases complete; 1 unresolved dilemma blocking further progress

---

TL;DR

Phase 9 (ai_client Batch A — 8 BC sites migrated) followed the plan's narrowing pattern (except Exception → except (SpecificType)). Six of the eight sites were subsequently re-classified by the audit as UNCLEAR — a state the plan did not anticipate.

The plan's anti-sliming protocol says "do not change the audit heuristic" but the heuristic does not recognize valid drain-body patterns (return ErrorInfo, set empty default, build err_item dict). The 6 sites have legitimate sinks; the audit just doesn't know about them.

Two options are evaluated below. Tier 1 decision needed before proceeding with Phase 10.

---

What was supposed to happen

Per conductor/tracks/result_migration_baseline_cleanup_20260620/plan.md:

• Phase 9 — ai_client Batch A: 8 INTERNAL_BROAD_CATCH sites (lines 332, 355, 394, 520, 537, 716, 723, 994)
• Phase 10 — ai_client Batch B: 8 more BC sites (lines 1528, 1599, 1611, 1636, 1657, 1854, 2848, 2867, 2898 — note: count is 9)
• Phase 11 — ai_client silent-swallow (9 sites): CRITICAL anti-sliming
• Phase 12 — ai_client rethrow classification (7 sites): Pattern 1/2/3
• Phase 13 — rag_engine migration (9 sites)

What actually happened

Category	Plan expected post-Phase 9	Actual post-Phase 9	Delta
INTERNAL_BROAD_CATCH (BC)	17 → 9 (-8)	17 → 9 (-8)	OK
INTERNAL_SILENT_SWALLOW (SS)	9 (unchanged)	9 → 11 (+2)	+2 from narrowing (set_tool_preset, set_bias_profile)
INTERNAL_RETHROW	7 (unchanged)	7 (unchanged)	OK
UNCLEAR	0 (not in plan)	0 → 6 (+6)	NEW GAP

The 6 UNCLEAR sites

Line	Function	Pattern	Drain
L332	_classify_deepseek_error	except (ValueError, AttributeError): → assigns body to fallback	Returns ErrorInfo (canonical drain)
L355	_classify_minimax_error	except (ValueError, AttributeError): → assigns body to fallback	Returns ErrorInfo (canonical drain)
L394	set_provider	except (OSError, ValueError): → fallback to empty api_key	Empty api_key call (safe default)
L716	_execute_tool_calls_concurrently (deepseek)	except (ValueError, TypeError): args = {}	Empty dict (safe default for malformed JSON)
L723	_execute_tool_calls_concurrently (minimax)	except (ValueError, TypeError): args = {}	Empty dict (safe default)
L994	_reread_file_items	except (OSError, UnicodeDecodeError) as e: → builds err_item	err_item["error"] = True (in-band error flag)

All 6 have legitimate drain mechanisms. None of them are silent-swallow (they propagate the failure to a structured destination — ErrorInfo, err_item dict, or empty default). The audit's existing heuristics don't cover these patterns.

Why this is a dilemma

The plan is self-contradictory in this area:

• (e) Anti-sliming protocol says "do not change scripts/audit_exception_handling.py" and "the audit heuristic is correct"
• (f) Classify-as-suspicious laundering is forbidden

But:

• The heuristic does not recognize the 6 valid drain patterns above
• Without heuristic coverage, the only way to silence the audit is either:
  1. Add a heuristic that recognizes the pattern, OR
  2. Migrate the site to a pattern the heuristic recognizes (e.g. return Result(...))

The previous sub-tracks (gui_2_20260619) handled this exact case in Phase 11 (dunder-raise heuristic) and Phase 12 (lazy-loading fallback heuristic). This sub-track's plan acknowledges those precedents but does not include equivalent heuristics for the new patterns.

Impact on remaining phases

If this dilemma is unresolved, the same pattern will repeat in Phase 10 (Batch B has 9 BC sites that will likely produce more narrow+fallback patterns → more UNCLEAR sites). Each subsequent phase risks:

• Plan-undercounted SS sites (currently +2 over plan)
• Plan-not-mentioned UNCLEAR sites (currently +6 over plan)

The plan's invariant tests assert:

• phase_11_invariant_ai_client_silent_swallow_zero (plan's stated target)
• phase_13_invariant_rag_engine_total_migration_target_zero

These assertions are based on the original baseline counts (9 SS, 0 UNCLEAR in ai_client). If we don't address the new sites, the assertions will fail or the audit gate will fail at Phase 14.

Options

Option A: Add audit heuristics (recommended)

Add 1-2 new heuristics to scripts/audit_exception_handling.py that recognize the 6 valid drain patterns:

1. Heuristic E: narrow-catch + drain-body — except (NarrowType): where the immediately-following body is one of:

   • return ErrorInfo(...) or return Result(errors=[...])
   • body = <fallback_value> where fallback is a documented safe default (empty dict, empty string, etc.)
   • <item>["error"] = True (in-band error flag pattern)
   • Build an err_item dict with error: True field

   This is the same approach sub-track 4 used for dunder-raise (Phase 11) and lazy-loading fallback (Phase 12). The plan acknowledges those precedents.

Pros:

• Honest classification of what's actually there
• 1-2 small heuristic additions, each with regression test in tests/test_audit_heuristics.py
• Future phases (10-13) don't need special handling
• Audit gate at Phase 14 will pass cleanly

Cons:

• Contradicts the "do not change the audit" instruction in plan §4 (but the contradiction is acknowledged as a plan bug)
• Requires 5-10 minutes to add heuristics + tests
• Sets a precedent that the audit can be amended mid-track

Option B: Full Result[T] migration for the 6 sites

Convert each of the 6 sites to return Result[T] with the fallback case propagated through Result:

def _classify_deepseek_error_result(exc, source) -> Result[ErrorInfo]:
    try:
        err_data = exc.response.json()
        ...
    except (ValueError, AttributeError) as e:
        return Result(
            data=ErrorInfo(kind=ErrorKind.UNKNOWN, message=exc.response.text, source=source, original=exc),
            errors=[ErrorInfo(kind=ErrorKind.INTERNAL, message=str(e), source=..., original=e)],
        )

Plus callers (_send_deepseek etc.) need updating.

Pros:

• Most "correct" per the styleguide
• Strictly Result[T] propagation as the convention requires

Cons:

• 6 call-site rewrites (or 6 _result helpers + 6 legacy delegations)
• Risk of breaking ai_client call patterns that rely on the current return shape
• Higher chance of test regression
• 30-60 minutes of work + test verification
• Doesn't actually solve the plan-not-anticipating-the-pattern problem — Phase 10 will likely produce MORE of these sites

Option C: Document and defer

Add a notes.md to the track that acknowledges the +6 UNCLEAR sites as a known gap, and adjust Phase 11's plan to include them. Don't fix the audit; don't migrate the sites. Phase 11 will need to add the heuristic OR migrate them then.

Pros:

• Minimal action now
• Tier 1 can evaluate and direct

Cons:

• Doesn't actually resolve the dilemma; same work happens later
• Phases 10-13 will keep producing more UNCLEAR sites

Recommendation

Option A. The pattern is small, well-defined, and precedent (sub-track 4 phases 11 and 12 added similar heuristics). It is the lowest-risk, fastest, and most consistent-with-prior-sub-tracks path forward. Phase 10-13 can proceed without special-case handling because the heuristic catches the pattern in all 3 baseline files.

What Tier 1 needs to decide

1. Approve Option A (add 1-2 heuristics to scripts/audit_exception_handling.py) — Tier 2 will proceed with Phase 10 after implementation
2. Approve Option B (full Result[T] migration of 6 sites) — Tier 2 will need ~30-60 minutes extra per Phase 10 site that exhibits the pattern
3. Approve Option C (defer to Phase 11) — Tier 2 continues Phase 10 with the caveat that the SS/UNCLEAR counts will diverge from plan
4. Other — Tier 1 may have a preferred approach not listed here

Current state of the branch

• Branch: tier2/result_migration_baseline_cleanup_20260620
• Last commit: 9a49a5ee (Phase 9 checkpoint)
• Commits ahead of origin/master: 50+
• Tests passing: 28 (Phase 1-9 invariants)
• src/mcp_client.py: 100% migrated (0 sites)
• src/ai_client.py: 24% migrated (8 of 33 sites; 6 NEW UNCLEAR sites added)
• src/rag_engine.py: 0% migrated (pending Phase 13)

Files for reference

• conductor/tracks/result_migration_baseline_cleanup_20260620/spec.md — design intent
• conductor/tracks/result_migration_baseline_cleanup_20260620/plan.md — executable plan
• conductor/tracks/result_migration_baseline_cleanup_20260620/state.toml — task status
• scripts/audit_exception_handling.py — the audit heuristic in question
• tests/test_audit_heuristics.py — 8 regression tests for the audit (precedent: 2 added in sub-track 4 Phase 11, 3 added in sub-track 4 Phase 12)
• docs/reports/TRACK_COMPLETION_tier2_autonomous_sandbox_20260616.md — sandbox convention reference
• docs/reports/TRACK_COMPLETION_result_migration_gui_2_20260619.md — most recent sub-track precedent

---

Awaiting Tier 1 decision before proceeding with Phase 10.