|
| 1 | +# P-014 PROD0 (coverage only) — report |
| 2 | + |
| 3 | +**Verdict: `PASS-P014-PROD0`.** G0–G3 hold (§5). |
| 4 | + |
| 5 | +What the PASS covers: an **opt-in, observation-only** extractor flag, `--binding-coverage FILE`. |
| 6 | +- **No default behaviour changes.** No OwnIR change, no finding, no SARIF and no exit-code change. |
| 7 | +- **PROD1 is not authorised.** PROD1 would feed the new symbols to diagnostics. |
| 8 | + |
| 9 | +Rules: [`00-preregistration.md`](00-preregistration.md), commit `6ec551f` (12:36:03 +05; the "12:40" in its header |
| 10 | +is my clock reading, the commit time is authoritative). Base `main` `22b29d5b`. Work 12:35–12:50 +05, within |
| 11 | +the 0.5 agent-day budget. |
| 12 | + |
| 13 | +Labels: |
| 14 | +- **[run]** measured here; |
| 15 | +- **[read]** checked in code or diagnostics; |
| 16 | +- **[doc]** dotnet/sdk source. |
| 17 | + |
| 18 | +## 1. What changed in production code |
| 19 | + |
| 20 | +| File | Change | |
| 21 | +|---|---| |
| 22 | +| `frontend/roslyn/OwnSharp.Extractor/BindingCoverage.cs` (new) | the #422 instrument's resolver and coverage code, moved into namespace `OwnSharp.BindingCoverage` with public types, plus the sidecar `Runner`. Behavioural differences from #422 are listed below | |
| 23 | +| `frontend/roslyn/OwnSharp.Extractor/Program.cs` (+17 lines) | two flags (`--binding-coverage FILE`, `--binding-coverage-tfm TFM`), usage text, and one call **after** the facts and the H0 file are written. A failure is reported on stderr and **never changes the exit code** | |
| 24 | +| `.github/workflows/ci.yml` (one step in `wpf-extractor`) | facts are `cmp`-identical with and without the flag, for project and directory input; directory and unrestored-project inputs are recorded `UNSUPPORTED`; a restored copy of `project-input-sample` is measured, deterministic, `foreign_binding = 0`, and COMPLETE exactly when no reason is present | |
| 25 | +| `frontend/roslyn/README.md`, `docs/proposals/P-014-semantic-resolution.md` | the flag documented; PROD1 marked as not authorised | |
| 26 | + |
| 27 | +**Differences from the #422 instrument.** All are stricter or needed to run on other hosts: |
| 28 | +1. **A project-reference member that is not `COMPLETE` makes its parent `SOURCE_SET_INCOMPLETE`.** In #422 only an |
| 29 | + unsupported or reference-missing member did. |
| 30 | +2. **A package with MSBuild `build`/`buildMultiTargeting` assets makes the unit `SOURCE_SET_INCOMPLETE`** |
| 31 | + (`package-msbuild-assets:<id>/<ver>`). Such assets can change the compilation. |
| 32 | +3. **The dotnet root comes from `DOTNET_ROOT`**, or else from the running runtime's location, instead of a |
| 33 | + hard-coded Windows path. The packages folder comes from `NUGET_PACKAGES`, or else `~/.nuget/packages`. |
| 34 | +4. **A targeting pack recorded in `downloadDependencies` is looked up in `dotnet/packs`, then in the packages |
| 35 | + folder**, which is where the SDK puts a downloaded pack [doc, `ResolveTargetingPack` behaviour]. Without this, a |
| 36 | + restored project on a host whose packs lack the exact patch would be a false `REFERENCE_MISSING`. That is fail |
| 37 | + closed, but wrong. |
| 38 | + |
| 39 | +## 2. G0 — package-vs-package conflict `Wpf.Ui.dll` against the SDK oracle [run] |
| 40 | + |
| 41 | +**Fixture.** `fixtures/G0`, my own project, `net10.0-windows` with WPF, referencing `unchihugo.WPF-UI` 4.4.2 and |
| 42 | +`WPF-UI.Tray` 4.3.0, restored offline from the global packages folder only. This reproduces P5's conflict: |
| 43 | +`unchihugo.WPF-UI/4.4.2`, AssemblyVersion 4.4.1.0 and PublicKeyToken `null`, against the transitive `WPF-UI/4.3.0`, |
| 44 | +AssemblyVersion 4.3.0.0 and PublicKeyToken `11f9f5cc97b3ffd6`. |
| 45 | +- **The oracle** chose `unchihugo.wpf-ui/4.4.2/lib/net10.0-windows7.0/Wpf.Ui.dll`. |
| 46 | +- **The #422 instrument** and **the production flag** chose the same file, by rule `assembly-version`. |
| 47 | +- **The full manifests are identical:** 209 = 209 (`data/G0.instrument422.compare.json`, `data/G0.compare.json`). |
| 48 | +- **No narrowing was needed.** |
| 49 | +- **Scope.** This qualifies one package-vs-package form: different AssemblyVersions, decided by the second rule. |
| 50 | + Package-vs-package conflicts with equal versions are still `CONFLICT_UNRESOLVED` by construction, and are not |
| 51 | + oracle-checked. |
| 52 | + |
| 53 | +## 3. G1 — P4/P5 are not reported as fully checked [run, `data/P*.cost.1.sidecar.json`] |
| 54 | + |
| 55 | +All 5 units, the 4 of P4 and P5, have reference status `COMPLETE`, `PACK_MAJOR_MINOR`, and |
| 56 | +**`unit_status = SOURCE_SET_INCOMPLETE`**, with named reasons. No unit is `COMPLETE`. |
| 57 | +- **Site counts and manifests** are identical to #422's for every unit. |
| 58 | +- **The only difference** is the added fail-closed reasons from §1: |
| 59 | + - `package-msbuild-assets:*`, e.g. `communitytoolkit.mvvm/8.4.2` in P5; |
| 60 | + - `p2p-unit-source_set_incomplete:*` in P4. |
| 61 | + |
| 62 | +## 4. G2 — compatibility [run, `data/run_prod0.log`] |
| 63 | + |
| 64 | +| Check | Result | |
| 65 | +|---|---| |
| 66 | +| Facts byte-identity: P4, P5 (`.csproj`, `--flow-locals --no-project-refs`) and `frontend/roslyn/samples` (directory) | **identical** across `main` 22b29d5b, this branch without the flag, and this branch with it (P4 `AD528B8A…`, P5 `5F0EA76D…`, samples `F82A32D8…`) | |
| 67 | +| Facts with the flag on and off, 6 runs per project | identical | |
| 68 | +| Directory input | sidecar `UNSUPPORTED (input-kind:directory)` | |
| 69 | +| In-memory controls against the **production** code (harness references the built `ownsharp-extract.dll`) | K1, K2, K4 (CS0101 merged; CS0433 at the site, separately compiled), K5, K6, K7, F2.3, F2.4, F2.5, K11, K12 and the guard: **all PASS**, same observations as #422 | |
| 70 | +| Fixtures through the production flag vs the SDK oracle | K8a 163 = 163, K8b 163 = 163, K8c 206 = 206, G0 209 = 209, K9 net8.0 163 = 163, K9 net10.0 167 = 167; no tie assumption. K9 without TFM: `UNSUPPORTED(multi-tfm)`. K10: with child assets, two `COMPLETE` units; without, `SOURCE_SET_INCOMPLETE(p2p-no-assets:Child)`, event unresolved, foreign = 0 | |
| 71 | +| `foreign_binding` | 0 in every control and in all P4/P5 units | |
| 72 | +| Sidecar determinism | byte-identical across 3 runs for P4 and P5 | |
| 73 | +| Build | only warning, before and after: the existing `Program.cs` CS8620, shifted 10 lines. No new warning | |
| 74 | +| `ruff check .` | passes | |
| 75 | +| `python tests/run_tests.py` | **not complete on this host.** It stops at `test_checkpoint_status`, which runs `cargo metadata`, and there is no `cargo` here. Every module before it passed. No Python file is changed by this branch; CI runs the full suite | |
| 76 | +| New CI step, rehearsed locally (offline restore) | facts identical; `UNSUPPORTED` for the unrestored and directory inputs; restored sample `COMPLETE`, `PACK_MAJOR_MINOR`, event bound, deterministic | |
| 77 | + |
| 78 | +## 5. G3 — cost [run] |
| 79 | + |
| 80 | +Medians of 3 separate processes (`data/run_prod0.log`, `data/P*.legacy.*.timing.json`). |
| 81 | + |
| 82 | +| | flag off | flag on | added by the flag | limit: 3 × legacy-merged warm (same session) | peak working set, on | |
| 83 | +|---|---|---|---|---|---| |
| 84 | +| P4 | 1103 ms | 2944 ms | 1841 ms | 3 × 1783 = 5349 ms | about 258 MB | |
| 85 | +| **P5** | 2844 ms | 4234 ms | **1390 ms** | **3 × 1021 = 3063 ms** | **about 302 MB** (limit 2.0 GB) | |
| 86 | + |
| 87 | +**Verdict.** G0 PASS (no narrowing needed), G1 PASS, G2 PASS, G3 PASS: **`PASS-P014-PROD0`.** |
| 88 | + |
| 89 | +## 6. Limits that stay attached |
| 90 | + |
| 91 | +1. **A complete reference set is not a complete compilation.** Real projects come out `SOURCE_SET_INCOMPLETE` |
| 92 | + (implicit usings, generators, XAML, package MSBuild assets). The coverage numbers measure binding under that |
| 93 | + incomplete source set; they are not a share of verified code. |
| 94 | +2. **The conflict resolver is qualified on these forms:** |
| 95 | + - framework vs package, decided by override, AssemblyVersion or FileVersion; |
| 96 | + - targeting-pack profiles; |
| 97 | + - target selection; |
| 98 | + - package vs package with **different** AssemblyVersions (G0). |
| 99 | + |
| 100 | + Not qualified: equal-version package-vs-package conflicts (left `CONFLICT_UNRESOLVED`), the rank and platform |
| 101 | + steps, and the full SDK contract. |
| 102 | +3. **Environment.** The host .NET is SDK 11.0.100-rc.2 with packs 8.0.32/10.0.13. The #419 packs (8.0.31/10.0.12) |
| 103 | + are gone, as recorded in #422. |
| 104 | +4. **Configuration outside the repository.** Any `Directory.Build.*` above a project, including a host file, is |
| 105 | + reported and makes the unit incomplete. It is not evaluated, and there is no guarantee about it. |
| 106 | +5. **The Python suite was cut short locally** by the missing `cargo` (§4). CI is authoritative. |
| 107 | + |
| 108 | +## 7. Files |
| 109 | + |
| 110 | +- `00-preregistration.md`, `REPORT.md` |
| 111 | +- `fixtures/`: G0, plus copies of #422's K8a, K8b, K8c, K9, K10 |
| 112 | +- `harness/`: the in-memory controls, compiled against the built extractor |
| 113 | +- `scripts/run_prod0.ps1`: build main and branch, facts identity, controls, oracle fixtures, determinism, cost |
| 114 | +- `data/`: control results, oracle comparisons, sidecars, timing, log. Local paths are replaced by `<nuget>`, |
| 115 | + `<dotnet>`, `<research>` and `<work>`. |
0 commit comments