swayfx-mosaic/docs/design/STEP20C_STATIC_RUNTIME_FOCUS_INPUT_STANDARD_PROBE.md

STEP20C static runtime focus, input and standard-return probe

EXPERIMENTAL / INCONCLUSIVE. Historical campaign frozen as evidence. Phase A published 5e2961d05b7b33252718aabc4dbf787724f79ef0 to both remotes, with a clean worktree. The original phase B was local and unpublished; the later evidence freeze publishes its sources and findings without replay. Architecture, UX, Packaging and Roadmap remain authoritative. STEP20A DONE, STEP20B EXPERIMENTAL, STEP20C NEXT, STEP20D–G PLANNED; production multi-output NOT YET SUPPORTED. No production, algorithm, shader, package or UX change.

Historical evidence and exact candidate

The previous runtime remains PARTIALLY_FEASIBLE / UNDECIDED. It captured h2-reference.png while workspace 2 was focused, then h2-mosaic.png after workspace 1/client setup and Mosaic. The full RGB comparison failed by 23,807 pixels, bounding box (0,0,800,600). Changes outside the unchanged reference interior support focus/title/border variation. No target colors appeared on H2; 182,214 survived on H1. Neither cropping borders nor a new successful run can retroactively pass that comparison.

Reuse only the historical staged binary from /home/fy59/swayfx-mosaic-static-runtime-hm5kbdkt/stage-static/usr/bin/swayfx-mosaic: SHA-256 f82557c1a9978a5e9665b9b4291b74880253725359e3df1c5b8d554921f7603b, ELF build ID 942db64269640778ee433e5970ba1100bc60f94c. Verify original/copy hash, executable regular-file mode, ELF dependencies and system wlroots SHA-256 7edca9a7d65855addda68e29ba5db94d57c803eb5fa7d0f86db3136b90ca9837. No SceneFX DT_NEEDED or loaded SceneFX shared mapping is allowed. Stop before launch if identity is unavailable. No SceneFX/Mosaic build or binary alteration.

Historical C17 saved-client and static-reference plus the extent-adapted virtual pointer are reused from manifest-verified source. Their historical executable files are not entries of those manifests: rebuild only these clients from the verified C/generated-protocol sources, record commands and resulting hashes. No client algorithm changes. Target alternates two opaque colors on configure; reference remains RGB (48,48,56) across focus/configure changes. Pointer uses 1600×600 logical extent and global-x + 800 normalization.

Single-run protocol and stop gates

One new private LAB, mode-0700 runtime, explicit private Wayland display and IPC socket, WLR_BACKENDS=headless. Every swaymsg uses -s. Only owned process groups are stopped; preserve exit codes and verify no residual process/socket. A 150-second deadline and bounded logs apply. No pristine run, no popup. HEADLESS-1=(-800,0,800,600), HEADLESS-2=(0,0,800,600), scale 1, normal.

Create reference on workspace 2, then filler and target on workspace 1. Both reference and filler have org.mosaic.StaticReference App ID. Therefore do not use the broad for_window App ID rule: obtain reference con_id while it is the only reference and apply [con_id=ID] border none in private IPC. Check reference PID, workspace/output, border none and zero border width; filler keeps its normal border. Keep focus_follows_mouse no. No daily configuration changes.

Activate Mosaic, read real applied owner/tiling traces and IPC scroll. Compute one signed camera delta from horizontal content offset and configured width, then reread scene/scroll. The historical +160 is not a constant. No further camera command or hide/return. Focus must initially remain target/workspace 1. If later normal focus changes move the camera, retain that observation and stop any assertion needing unchanged Mosaic geometry; do not correct it.

PI_OWNER gives an applied scene owner, not a buffer. IPC rect/window_rect/deco_rect and the client's last real COMMIT size are separate evidence. Use the actual H1 painted color rectangle to identify vertical content and pick an interior point. Correlate measured paint with horizontal scene offset and client width; never label the old source_box y estimate a native buffer coordinate. PI_SCENE local clip plus output-parent origin and tiled translation gives global viewport.

After complete Mosaic setup and sole calibration, park pointer at (-790,590), wait for stabilization, capture H2 twice at least 500 ms apart. Check 800×600, full-frame reference color, border, focus/workspace and exact whole RGB equality. Store PNG hashes and snapshots. No ROI, no tolerance. Failure stops before input.

V1: capture H1/H2, retain positive target-color count on H1, zero on H2, and whole-H2 equality to preflight. Require content crossing x=0 and clip ending there. Save every differing pixel coordinate/color on failure; stop dependent tests.

V2: choose a point at least 20 pixels inside the actually painted H1 rectangle, outside decorations and borders. One right press/release; require actual target BUTTON 273 states 1 then 0 with recorded serials, no reference delivery. A move alone is never input proof. Preserve available passive scene hit and Wayland logs.

V3: choose an H2 point within the geometrical continuation of the target content and the full-frame reference interior. One right press/release; require zero new target BUTTON events and two real reference events. Record focus after click, park pointer, stabilize and recapture whole H2 with exact equality. Do not infer PASS from neither client receiving input.

V4: address the target con_id with layout splith (after V3 global focus is H2; a bare layout command could affect the wrong workspace). Verify workspace identity, applied splith, target still mapped, applied owner and PI_SCENE clip_valid=0 on H1 and ordinary H2. Capture H1/H2. Choose a fresh interior target-colored point from the standard image, inject one click and require actual target press/release. Park pointer and recapture H2 after focus transition. Require stable H2 and no residual clip. Do not reuse the Mosaic click point blindly.

Budget, analysis and classification

Persistent ledger, exclusive flock, atomic durable reservation before run, each PNG, each click, camera and bounded pointer movement. Failed reservations remain consumed; no reset/refund or alternate LAB. Offline tests must precede run. Limits: run 1, pristine 0, SceneFX/Mosaic builds 0, camera adjustment 1, each of three named clicks 1, PNG 12, pointer moves 8, duration 150 seconds. Expected PNG count 8. All generated artifacts stay outside Git. Seal source and evidence manifests once after collecting results, then verify every entry without rewriting.

Keep client damage requests, scene/output damage extents and stabilized PNGs separate. No claim of complete intermediate frames or native output-buffer/PNG identity. Root/nested grabs, other geometry, migration, saved buffers/effects, full semantic ABI, other wlroots pairs, legal notice closure and named release/ security owners are NOT TESTED or remain open. D1–D4 stay approved privately.

FEASIBLE requires every V1–V4 assertion and conformant binary/dependencies, no crash or blocking defect in this matrix. PARTIALLY_FEASIBLE retains demonstrated principal properties with unqualified input/return/pixels. INCONCLUSIVE means an unreliable fixture. BLOCKED requires demonstrated pixel/input leakage, runtime ABI defect or standard regression. Focus/calibration error alone is not a clipping failure. Architecture: STATIC_MOSAIC_FOCUS_INPUT_STANDARD_FEASIBLE_FOR_TESTED_MATRIX, STATIC_MOSAIC_FOCUS_INPUT_STANDARD_REQUIRES_REDESIGN, or UNDECIDED respectively.

Independent future reconstruction (not authorized in this campaign)

A later campaign can archive pinned Mosaic cd550382b3, SceneFX 37ccd723bef49e6891156ffafce8f549f01446cc and wlroots d783533489e1f75d6886c2ab5c5960090ef268f8, without the old LAB. Apply the ordered three SceneFX and five Mosaic patches listed in the static distribution report. Use SceneFX default_library=static/examples=false, its uninstalled pkg-config provider and Mosaic --prefix=/usr --bindir=bin with mosaic-package=true, swaybar/swaynag=false, man-pages=disabled, default-wallpaper=false, sd-bus-provider=libsystemd, werror=false. Build in independent directories, qualify exact shared wlroots, model, ELF, notices/stage and new authorized runtime budgets. Preserve toolchain/generated protocols/patch hashes. This is source reconstruction, not a promise of byte-reproducibility or permission to substitute another binary in this ticket.

The fixture provides the bounded runner, offline tests and evidence collector.

Completed campaign — retained fixture failure

SWAYFX_MOSAIC_STATIC_FOCUS_INPUT_STANDARD_PROBE=INCONCLUSIVE

Architecture: UNDECIDED. LAB: /home/fy59/swayfx-mosaic-static-followup-nhwejz6u. One run, 2.661722 seconds. The unique runtime stopped before the first screenshot, movement or button. No restart, second LAB, binary/algorithm repair or budget refund occurred.

Identity, setup and measured state

The copied executable retains SHA-256 f82557c1a9978a5e9665b9b4291b74880253725359e3df1c5b8d554921f7603b, ELF build ID 942db64269640778ee433e5970ba1100bc60f94c, mode 0755. Readelf and ldd -r pass; process maps resolve the qualified system wlroots 0.20.2 binary with SHA-256 7edca9a7d65855addda68e29ba5db94d57c803eb5fa7d0f86db3136b90ca9837. No shared SceneFX or historical LAB dependency appears in the runtime mappings. Meson/model evidence is reused from A; no new compositor/library build.

Three C17 client-only builds from manifest-verified sources passed syntax and compilation and produced the same bytes as the historical executables:

Client	SHA-256
saved-client	3d5d0c5e446b995c210006a5bafe0d86dd5f32a228e1b7137d12af6ec9e217da
static-reference	11166ea1bb7cb3bb08f10ab93e091c5c5ea8687649b5fb0705288c7ca91680e8
pointer-daemon	76ad0f093f2b7a97359920b7bd456cfa93a1581c73875547a86eb911f2ed2d7a

These hashes are newly recorded; the historical executable files themselves were not sealed. Historical C/generated protocol inputs were sealed and verified. The new retained source/command/binary chain supplies explicit provenance.

Both outputs matched the requested negative-origin 800×600/scale-1/normal matrix. Private configuration retained focus_follows_mouse no. The command [con_id=7] border none succeeded for reference PID 1674352 on workspace 2 (ID 5), H2. Filler ID 8, with the same App ID, kept normal border and width 2. Target ID 9/PID 1674361 on workspace 1 (ID 6), H1 remained focused; reference was unfocused. No broad App ID rule affected the filler.

Applied camera read 0 before calibration. The target's horizontal content offset was -158 with configured width 636; desired left -318 gave signed delta +160. IPC and scene both read 160 afterward. Applied owner scene became (-320,0), size 640×600. IPC window offset x=2 and actual client COMMIT size 636×573 support horizontal content continuation [-318,318), crossing x=0. They do not prove native buffer y or actually visible pixels without a capture. PI_SCENE records tile-local (-160,0), clip (160,0,800,600), translated through H1 origin (-800,0) to global clip (-800,0,800,600). H2 records clip_valid=0. This proves the observed applied clip state, not the unperformed pixel/input gates.

Exact stopping assertion and source diagnosis

The executed runner required both border none and current_border_width==0. Reference IPC instead reports:

{"border":"none","current_border_width":2,
 "rect":{"x":0,"y":0,"width":800,"height":600},
 "deco_rect":{"x":0,"y":0,"width":0,"height":0},
 "window_rect":{"x":0,"y":0,"width":800,"height":600}}

This failed the stored-width predicate before writing owner-geometry.json or reference-preflight.json and before any capture reservation. Offline collection reconstructs the actual owner/reference observations from retained IPC/traces and explicitly labels all unexecuted results NOT TESTED.

Source diagnosis: sway/commands/border.c:cmd_border calls set_border(B_NONE); it changes border_thickness only with a second argument. sway/ipc-json.c publishes current.border_thickness unconditionally. sway/desktop/transaction.c sets the effective border width/top to zero for B_NONE. Therefore stored width 2 is compatible with no border. The runner confused stored configuration with effective decoration geometry. No rendered border or Mosaic leak is demonstrated. The initial protocol's zero-width expectation above is retained as the failed hypothesis; it is not a new product contract. The exact executed runner remains unchanged, and no corrected runner is represented as executed.

Matrix, damage and cleanup

Gate	Result
Binary, loader, exact wlroots	PASS
Output geometry, client mapping, focus, targeted border mode	PASS IPC/source evidence
Calculated camera and applied clip	OBSERVED; no pixel qualification
Two H2 reference captures / whole RGB equality	NOT TESTED; zero PNG
H1 retained pixels / H2 forbidden colors	NOT TESTED in B
Visible Mosaic button	NOT TESTED
Invisible Mosaic non-delivery / reference H2 delivery	NOT TESTED
H2 focus-transition RGB stability	NOT TESTED
splith return / clip removal / standard button	NOT TESTED
Root/nested popup or grab	NOT TESTED, deliberately outside scope
Damage	Setup/client/scene/output extents retained; no PNG correlation or intermediate-frame coverage
Crash/protocol error	No crash; each ordinary client exited 0 with error=0
Cleanup	PASS; compositor exit 0, three clients exit 0, owned pointer terminated -15; PIDs and private runtime absent

The owned runtime was /tmp/msr-330t08c8, removed after all owned process groups were reaped. No daily IPC, installation, system dependency replacement, package, production patch/source, UX, dotfile, QuickShell, Codex or STEP20D change occurred.

Budget: SceneFX build 0, Mosaic build 0, binary modification 0, run 1/1, pristine 0, camera 1/1, captures 0/12, each named click 0/1, moves 0/8. The consumed run prevents further execution despite unused PNG/click capacity. The persistent owner registration and ledger are retained and never reset.

The future smallest justified gate is a separately authorized single run after correcting the fixture's border-semantic predicate and testing it offline against this retained border-none/width-2 snapshot. Keep the same focus/RGB/input/splith objectives and exact candidate identity if still available. No SceneFX or clip algorithm modification is justified by this fixture failure. This ticket does not close the static runtime continuity gate or authorize STEP20C integration.

Q1–Q2 reference stability and full RGB: NOT TESTED. Q3 forbidden pixels: NOT TESTED in B; A's zero target colors remain historical only. Q4–Q9 button delivery, invisible input, focus RGB, clip removal and standard input: NOT TESTED. Q10 static/shared behavioral equivalence: not established by this run. Q11 algorithm change justified: no demonstrated runtime anomaly. Q12 wider geometry, lifecycle, full damage/effects/saved buffers, applied-generation and exhaustive ABI remain unqualified, as do this run's intended principal gates. Q13 preparation for integration: the static build/load candidate is established, but the requested continuity evidence remains open. Q14 next technical gate: correct the fixture and obtain a newly authorized bounded RGB/input/standard-return observation. D1–D4 persist; STEP20C NEXT, production multi-output NOT YET SUPPORTED.

Validation and integrity receipt

Python AST, shell/recipe syntax, concurrent/exhausted persistent-ledger tests, exact-RGB and actual-button offline tests, retained border snapshot/source semantics, local Markdown links, source-only fixture, protected source/system hashes, unchanged executed runner and owned-process/socket cleanup: PASS. Runtime scenario: FAIL at fixture assertion, scientific verdict INCONCLUSIVE. Meson/model: historical qualified results reused, NOT RUN anew. Standard test suite/headless campaign: NOT RUN, would exceed the authorized runtime scope. No subagent review was requested; the parent performed the bounded source, contract, provenance and final evidence audit. No independent review is claimed.

188 TEMPORAL_DAMAGE trace lines are retained from setup. They establish no post-click, standard-return or every-frame damage result.

B source manifest: 24 entries, SHA-256 5f9ab2b6d1ca23ef7743623d1e544237c8b70b62fe6de8164a46f1342e5a8ee2. B evidence manifest: 46 entries, SHA-256 41cc5e3d305d89e83314723e67d7116c36d1503dd35647894bc898c49343e1a5. Both were created once and every entry verified. Executed fixture and later collector/final fixture are distinct source directories. The report and final Git diff are outside these inventories to avoid circular digests. A's original manifests still verify without rewriting or snapshot remapping.

Final tracked B delta: this report plus Architecture, Known Limitations, Roadmap, documentation index and Changelog; source-only follow-up fixture added. No AGENTS, UX, Packaging, PKGBUILD, cumulative patch or production source change in B. Local main and both remotes remain the A publication commit 5e2961d05b7b33252718aabc4dbf787724f79ef0; B remains unstaged and uncommitted.

Static focus fixture evidence freeze — 2026-10-13

This new freeze task starts at main 5e2961d05b7b33252718aabc4dbf787724f79ef0. Direct git ls-remote queries to canonical Forgejo and GitHub both returned that exact commit. The expected local documentation and source-only fixture were audited in full. No foreign change was found. This preservation commit does not alter the historical executed runner, either historical manifest, or any raw observation. The earlier LOCAL/UNCOMMITTED statements describe the original campaign's state, not the publication state of this frozen evidence.

All 24 source and 46 evidence entries pass their original SHA-256 inventories; the manifest digests above are unchanged. The Python files match the sealed final-fixture byte for byte, and run.py matches the actually executed fixture. Python AST, offline RGB/button tests, concurrent/exhausted ledger tests, read-only collector verification, shell recipe syntax, local Markdown links and git diff --check pass. Production source, package recipe and cumulative patch are unchanged. No graphical test, compositor/library build or historical budget reservation is performed by this freeze. Run 1/1 and camera 1/1 remain consumed; unused captures/clicks/moves cannot authorize another historical run.

The source audit also reads the pristine pinned SwayFX border command: fd71a6bdc061bd633b488ae7b83e8a6981d22f86:sway/commands/border.c. Its B_NONE handling and optional second-argument thickness assignment match the retained candidate semantics. IPC thickness is stored metadata; effective border/decor geometry is separate. The failure remains INCONCLUSIVE / UNDECIDED. The separately authorized RGB/input/splith ticket must supply a corrected pure predicate, replay actual IPC and audit the entire preflight before its own run.

SWAYFX_MOSAIC_STATIC_FOCUS_FIXTURE_FREEZE=PASS