v3 prerequisites: fixes that distort the v3 measurements (§6)

[nikolay-e]

Engineering fixes that distort exactly the v3 measurements. All land before the dcbench split is frozen; the Q-class ones force a full calibration rerun if they land later. State verified at eb8f1a4f.

• 1. Truncated generated fragments priced at 0 — fixed in 0b787e09: fragmentation.rs prices truncated content + per-fragment overhead. Used-tokens / budget-compliance columns are honest again.
• 2. BM25 cell has no admission gate — scoring.rs BM25Scoring still builds with admissible_files: None (it builds no graph). 53ba7b64 compensates by raising τ to UNGATED_STOPPING_THRESHOLD (0.12) when τ is the default — but an explicitly passed lower τ in BM25 mode is still ungated. Decide for cell 3: either derive admission for BM25 from the discovery universe (the gate then changes this cell's numbers — hence "before, not after") or document cell 3 as ungated-by-construction. Q-class either way.
• 3. Empty-selection disclosure lost — fixed in 84bb6a0d / 0b787e09: every early exit returns ChangeSet::Empty { lockfile_changes, ignored_changes, policy_excluded_count }; regression test in tests/test_diffctx_invariants.py. C5's coverage block can rely on honest empty paths.
• 4. stand_in_ids over-wide / mis-scoped — settled in d9af5925 / 53ba7b64: every sig_lookup pairing, with the safety argument in select.rs; carries_change in types.rs consumes it. The narrower definition was tried and regressed the stub role.
• 5. memory_pipeline vs git.rs parity on pure deletions — PARTIAL: memory_pipeline.rs has a_pure_deletion_anchors_where_git_anchors_it, but its git side is a hand-written DiffHunk literal, not git::parse_diff output; nothing forces it to track git.rs. Needed: one test that runs both pipelines over the same deletion and asserts identical hunks (this is the "calib data comes from the same system" guarantee). Overlaps memory_pipeline re-spells the heavy phase: the #149 anti-drift fix covered only the selection half #232 (score_from_fragments shared prologue) — do them together.
• 6. GIL held in pybridge — PARTIAL: the six pipeline entry points release the GIL via detach_guarded; still held: build_project_graph (full walk + tree-sitter parse under the GIL — the standout), resolve_diff_range (git subprocess), count_tokens, and the graph analytics that borrow &PyProjectGraph. For honest harness wall-clock p50/p90, build_project_graph and resolve_diff_range must detach before the matrix run; the analytics can wait.

Also blocking budget-compliance claims (C4): #241 (--budget is not a hard cap — cores bypass it, 3.15× at 1000) and #238 (one generated file takes 88% of a 15-file range).

Gate for every item: scripts/bitcheck.sh for E-class, corpus + re-baseline for Q-class; then freeze.

[nikolay-e]

Four of six items closed in 8d582ef; the tracker is now:

• 1 — truncated generated fragments priced at 0 (already fixed in 0b787e0)
• 2 — BM25 has no admission gate: decided, not papered over. BM25 builds no graph, so it has no declared-related set to gate on; cell 3 is ungated by construction. The effective gate and tau are logged per run (selection gate=none|admission tau=…), because an ablation cell that cannot say whether it ran gated is not comparable with one that did. An explicitly requested tau is still honoured as given, so a sweep measuring an ungated tau gets the tau it asked for.
• 3 — empty-selection disclosure (already fixed in 84bb6a0 / 0b787e0)
• 4 — stand_in_ids (already settled in d9af592 / 53ba7b6)
• 5 — deletion parity: both_pipelines_anchor_a_pure_deletion_identically now compares diff_to_hunks against real git diff --no-index --unified=0 output through parse_hunks_from_diff_output. The old test compared against a hand-written DiffHunk literal, which could not notice git.rs drifting away from it.
• 6 — GIL: build_project_graph, resolve_diff_range and count_tokens now detach. The graph analytics that borrow &PyProjectGraph still hold it; they are not on the latency path the matrix measures, so they are deliberately left.

Also landed: #241 (envelope charged) and #238 (per-file share), both listed here as blockers of the budget-compliance column.

Not yet done, and the one thing still standing between here and the freeze: #233's full path — the 45 duplicate cells are gone, the smoke cell asserts and gets HF_TOKEN, and python -m eval equivalence can finally fail, but the Hetzner provisioning + 20 GB bake has still never executed, and results/sweep_v2_local/equiv/manifests exists on one laptop only. That is what #252's freeze depends on.