ed9375797e
Record exact-SHA scale and extended G samples, same-source blur measurements and retained negative saved/input results. Keep STEP20C NEXT and production multi-output unsupported; no production integration. Preserve immutable probe fixtures, offline regressions and compact artifact provenance. Fourteen consumed runs / 125 output captures, independent final audit accepted. Saved timeout/GPU lifetime, whole-H1 return, input after camera mutation and all-frame damage remain open.
88 lines
5.1 KiB
Markdown
88 lines
5.1 KiB
Markdown
# STEP20C targeted fixtures
|
|
|
|
EXPERIMENTAL / PLANNED. These source fixtures repair bounded measurement
|
|
procedures identified in the overnight evidence. STEP20C remains NEXT;
|
|
production multi-output remains NOT YET SUPPORTED. No production integration,
|
|
UX policy, package dependency, compositor build, installation or live socket
|
|
operation is part of this fixture.
|
|
|
|
The new private laboratory is
|
|
`/home/fy59/swayfx-mosaic-targeted-7gge6h8b` (owned mode 0700). The historical
|
|
`/home/fy59/swayfx-mosaic-overnight-mcolx0gq` is read only: its 18 runs are
|
|
exhausted. Historical verdicts remain unchanged. Public Python tests accept
|
|
`--historical` so paths can be supplied explicitly.
|
|
|
|
`plans.py` declares 13 accessible attempts: scale A/G, held saved A/G, effects
|
|
pristine/A/G, refreshed input pristine/A/G and G dynamic/hidden/R90. The
|
|
independent ledger permits at most 14 attempts, at most 16 compositor PNGs
|
|
per attempt, campaign PNG caps scale24/saved30/FX48/input36/G36, and total174.
|
|
Current planned allocation is164. The unused saved diagnostic is disabled;
|
|
it requires a separate parent decision, exact diagnostic identity and distinct
|
|
review/freeze. No diagnostic attempt is consumed by preparation.
|
|
|
|
Every launch/action reserves first in the same locked persistent ledger.
|
|
It uses `flock`, atomic replacement, file and directory `fsync`; a consumed
|
|
run token persists on failure. There is no reset, refund or retry operation.
|
|
Local click/motion/grab/chain/camera/mutation/configuration/spawn/capture limits
|
|
are reconciled from every reservation. The watchdog stops active measurement
|
|
at143 seconds and limits the total run logs to32MiB. Cleanup retains terminal
|
|
IPC response, normal process codes, rescue status and150-second bound.
|
|
|
|
`engine.Session(spec, auth, lab)` exposes `owner`, `find`, `cmd`, `client_cmd`,
|
|
`capture`, `pair_capture`, `click`, `close_client`, `debit`, `spawn`, `log`,
|
|
`launch_client`, `wait`, `save`, `calibrate` and `inject`. Clients are under
|
|
`LAB/clients`; raw run evidence is `LAB/runs/<id>/run`. All compositor/control
|
|
IPC explicitly names its private socket; inherited display/library selections
|
|
are removed.
|
|
|
|
Scale/R90 uses five marker pixels and the reviewed pre-owner layer-host
|
|
baseline before identifying one new XDG root. The same shared root validator
|
|
replays actual D2 and rejects host substitution, multiple candidates, missing
|
|
host, incorrect ancestry and incoherent callback order. Replay qualifies
|
|
association only; it does not invent missing historical input/final evidence.
|
|
|
|
Held saved uses source-audited `gaps inner current set 30` and startup
|
|
`-Dtxn-timeout=2500`. It requires mapped real SHM, a previous actual ACK,
|
|
a new received configure held without ACK, and a natural waiting transaction
|
|
timeout with the client alive. Actual argv plus source and elapsed timeout
|
|
corroborate2500ms; no native configured-value log exists. A/G expose no saved
|
|
node/render identity. SHM and pixels cannot fill that gap, so saved clipping
|
|
remains INCONCLUSIVE / SOURCE ONLY.
|
|
|
|
Input recomputes painted global coordinates after region restore, focus,
|
|
camera and layout mutations. It requires fresh local pointer coordinates,
|
|
actual client wire/button serial and target. The empty region intentionally
|
|
uses the previous global point to compare no owner delivery. Historical F
|
|
analysis processes ALL exact output/commit_seq pairs and reports missing
|
|
presents. No atomic generation, presented buffer identity or GPU lifetime is
|
|
inferred. Dynamic G samples fresh popup input after each resize/migration;
|
|
legitimate popup_done is retained and fresh toplevel input is sampled without
|
|
recreating a popup chain. Hidden G retains the same chain on return.
|
|
|
|
Preparation and validation are offline:
|
|
|
|
```sh
|
|
PYTHONDONTWRITEBYTECODE=1 python tooling/probes/step20c-targeted/prepare.py --lab LAB
|
|
PYTHONDONTWRITEBYTECODE=1 python tooling/probes/step20c-targeted/test_targeted.py --historical HIST --output LAB/offline.json
|
|
PYTHONDONTWRITEBYTECODE=1 python tooling/probes/step20c-targeted/replay.py --historical HIST --output LAB/replay.json
|
|
```
|
|
|
|
Preparation verifies retained binary/client/protocol hashes and privately
|
|
compiles only the new FX client. Generated protocol headers/C, binaries, SHM,
|
|
logs, captures and runtime directories remain outside Git. `identity.json`
|
|
records copied client source SHA and provenance closure. `gate.seal(LAB)`
|
|
requires14 passing core tests and5 passing FX tests, executed commands, raw output and exact source bindings, plus typed actual D2/F/E2 replay receipts and binds all fixture/import/client/tool/
|
|
binary/library/source evidence. A separate independent reviewer must write
|
|
`LAB/review.json` with scope `STEP20C_TARGETED_FINAL_BYTES_V1`, exact freeze SHA
|
|
and file bindings, accepted read-only independent review, empty blockers and
|
|
bound report bytes. The author does not create that acceptance.
|
|
|
|
Only after independent acceptance and explicit runtime authorization:
|
|
|
|
```sh
|
|
PYTHONDONTWRITEBYTECODE=1 python tooling/probes/step20c-targeted/run.py --lab LAB scale-A
|
|
```
|
|
|
|
Each invocation is one unique attempt. A missing or changed acceptance fails
|
|
before a run token, process or runtime exists. This fixture does not launch an
|
|
entire campaign automatically. Unexercised matrix cells remain NOT TESTED.
|