Pre-spec research — CSF #2670 landing-pass campsite follow-ups

Base: upstream/main @ c6b70d22e (research worktree research/csf-2670-followups, read-only). Squad: researcher-robbie (related-issue hunt), paula-patterns (SSOT/duplication root-cause), architect-alphonso (design decisions). All claims verified against current code, not ticket text alone.

Scope tickets (all OPEN, children of epic #1928 ruff/mypy/Sonar debt)

  • #2671 — recurring shard-map registration gap (design decision)
  • #2673 — bite-battery _SourceMutation xdist flake (isolation fix)
  • #2674 — sync remediation-guard escape + duplication (SSOT hardening)
  • #2675 — 17 pre-existing mypy errors / 8 files (type debt)

Root-cause findings (not error counts)

#2671 — one seam, two reds

A new tests/architectural/*.py trips both GC-1 completeness (test_arch_shard_marker_completeness.py:100) and the orphan ratchet (test_gate_coverage.py:571) from the same missing entry. Both resolve from one seam: shard_for() in tests/_shard_registry.py:129 (marker applied iff non-None). NOT a false-green — the arch-adversarial job runs if: always() with no path filter, so it is caught on the PR, but as an expensive confusing double-red needing a non-obvious manual table edit + re-push. A reminder-based fix has already failed 4×.

  • Recommendation (alphonso): Direction A, refined — per-group opt-in default_fallback field on ShardGroup; unassigned arch files route via a stable hash-bucket (sha1(relpath) % n + 1, NOT "lightest" — lightest piles all fallbacks on one shard); explicit _ARCH_SHARD_* entries still win; keep the GC-1 union invariant as the fallback's correctness net. One red-first WP. Seams: tests/_shard_registry.py:129 + tests/_arch_shard_map.py (opt-in flag + header-doctrine rewrite).
  • Ticket-text correction (robbie): it is 3 shard-map registration appends (test_marker_baseline.py #2590, test_workflow_dist_lint.py #2590, test_no_parity_scaffold.py #2651), plus a 4th golden-count annotation (c6b70d22e) which is a different mechanism. Spec should say "3 registration appends across ~3 incidents."

#2673 — shared mutable real-file state

_SourceMutation (test_single_mission_surface_resolver.py:822) mutates the real src/specify_cli/core/mission_creation.py on disk; a concurrent xdist worker's discover_rows() scanner reads it mid-mutation → false RED on an uninventoried _wp04_bite_witness sink. CI-safe today (injector shard_1, scanner shard_2 = separate checkouts); the flake is the local -n auto run only.

  • Recommendation (alphonso): Option (i) — mutate an isolated tmp copy + process-local monkeypatch of audit.SRC_ROOT; xdist workers are separate processes so the sibling scanner keeps reading the never-mutated real tree. Preserves what the battery proves (real detector code path, root-agnostic). Option (ii) serialize is blocked: xdist_group needs --dist loadgroup but repo mandates --dist loadfile. There are ~6 sibling _SourceMutation sites (a whack-a-mole class) — fix the #2673 pair now, leave a tracked note.
  • Duplicate (robbie): #2638 is a de-facto duplicate — a second scanner victim (test_surface_resolution_audit.py) of the same injection. One fix closes both; close #2638 as fixed-by.
  • Theme-twin (robbie): #2672 — same "tests must not mutate real repo files" hygiene root (pollutes real synthesis-manifest.yaml + ANSI). Cheap NO_COLOR=1+tmp-isolate fold candidate.

#2674 — one structural root (split surface + half-blind guard)

Guard test_no_unknown_commands_in_hints iterates only _REMEDIATION_HINTS dict (test_preflight_remediation_hints.py:118,182); the inline _build_remediation_lines() (preflight.py:307-334) emits orphan-daemons, sync migrate, auth login that are never validated to resolve under --help — a typo ships green. Plus daemon_server_url/team_or_user remedies are byte-identical duplicates across dict + inline (S1192 is semantic not lexical — different line-wrap points, so grep misses them).

  • Recommendation (paula): one registry — hoist all remedy prose to named constants; both the dict and the builder reference them; expose ALL_REMEDIATION_TEXTS; repoint the command-resolution guard at the registry (not the dict). Closes duplication + coverage in one seam. Small blast radius (one module + one test file). Whack-a-field trap: hoisting the 2 literals without widening the guard leaves orphan/migrate/auth-login unvalidated — incomplete.

#2675 — 11 target-file errors collapse to 6 roots

  • Cluster 1 (str→Lane, 4 errors @ workflow_executor 807/816/837/1232) → 1 root: get_wp_lane (status/lane_reader.py:50) returns Lane | str only because of legacy string sentinel LEGACY_UNINITIALIZED_SENTINEL. Fix-once: promote the sentinel to a Lane member (enum is StrEnum; Lane.GENESIS already models unseeded) → loader returns pure Lane, all 4 vanish, and get_all_wp_lanes annotation is corrected for free. Aligns with no-legacy-resolver-paths. (Pragmatic alt: Lane(...) coerce at 2 call sites 770/1209 — leaves the union leak.)
  • Cluster 2 (no-any-return, 2 @ 428/1372) → 1 root: _wf().locate_work_package leaks Any via ModuleType accessor. One typed wrapper _locate_wp(...) -> WorkPackage.
  • Cluster 3 (decision_id str|None→str, 2 @ specify_interview.py:181 / plan_interview.py:181) → 1 root replicated by duplication: narrowing captured into an intermediate bool mypy can't carry; the two files are byte-identical blocks. Consolidate into shared widen/interview_helpers.py + narrow once.
  • Cluster 4 (3 independent Optional narrowings @ workflow_executor 668/873, status_transition.py:854) → 3 roots, NO shared boundary. Local narrowing on a genuine invariant each. Do NOT invent a shared seam or blanket-cast.
  • Traps: blanket cast(Lane,...) / per-site cast(WorkPackage,...) / per-file decision_id narrowing all re-open the leak. Charter: fix the code, not suppress.
  • Other files (mechanical, 6 redundant-cast): _read_path_resolver.py:1473, mission_type_profiles.py:577/581/876, status/emit.py:302, m_2_1_4_enforce_command_file_state.py:78 — drop cast, watch unused import.

Collision with in-flight #2651 / mission-type-single-source-gate-wiring / S0

No production-code collision. Contention is test-surface only:

  • #2671 edits tests/_arch_shard_map.py — contended (#2651 already appends there; HEAD c6b70d22e is shard-map-adjacent; S0 adds arch tests). Landing #2671's fallback first removes the exact friction those missions keep hitting.
  • #2673 edits tests/architectural/test_single_mission_surface_resolver.py — this is the resolver-seam mission's PRIMARY test file (dense #2651 tags; S0 is SPECCED+PLANNED, still adding tests). Sharper collision — sequence #2673 for a quiescent window or fold into that mission's lane.
  • mission_type_profiles.py (6 redundant-cast in #2675) is owned by the mission-type work — coordinate those 3 cast-drops or defer them.
# Verdict Why
#2638 FOLD-IN → #2673 duplicate root (2nd scanner victim); close as fixed-by
#2672 FOLD (theme-twin of #2673) same real-file-mutation hygiene root; cheap
#2475 ADJACENT arch marker vacuous under .worktrees/ — different mechanism
#2476 ADJACENT/OUT proto-spec local arch-pole parity CLI — bigger, under #1931
#2607 ADJACENT GC-2b abs-path ids — different surface
#2625, #2631, #1843 OUT different surfaces / strategic
#1928 PARENT home epic for #2675 + cluster; link, don't scope

ONE quick-follow mission, theme "campsite-clean the #2670-landing residue so main stops going red for process reasons," with two rigour tiers:

  • Mechanical burn-down: #2674 (SSOT), #2675 (6 roots), #2673 (+#2638 dup, opt #2672).
  • One design decision: #2671 (Direction A) — its own decision-moment WP.

Open forks for the operator (below) before speccing.

Edit this page