Skip to content

Latest commit

 

History

History
229 lines (193 loc) · 42.6 KB

README.md

File metadata and controls

229 lines (193 loc) · 42.6 KB

Documentation Tests

Native child relationships: partial manifest, paired trees, and source-path traces recover contextual construction on symbol-rich 18.0.9246 arm64. just skin-child-check checks positive/negative controls and evidence integrity; just native-skin-children --parent button queries the findings. See scope and reproduction. This does not close the allowed-child vocabulary.

Native composition: object-tree capture and ABI evidence record class merging, source-order sensitivity, coordinate evaluation, deck propagation, and layout round trips in the 18.0.9246 arm64 custom host. just offline-skin-composition-check validates the saved paired cases; just offline-skin-composition show --case nested_relative prints a tree without vendor execution. These remain offline leads, not live-app parity or a complete schema.

Native skin host: capture, rendered frame, and routine evidence record a verified offscreen render on 18.0.9246 arm64. just offline-skin-host-check verifies all fixture pixels and retains the full-engine graphics-startup limitation.

Unknown-name follow-up: paired native results and hardware read-watch traces distinguish context-specific non-consumption from global name recognition on 18.0.9246 arm64. Validate with just offline-xml-unknowns-check.

Offline XML spike: native results, call/return traces, and layout evidence are pinned to 18.0.9246 arm64. just offline-xml-check validates saved evidence without executing vendor code. See scope and reproduction; these are offline leads, not live rendering proof.

Symbol-rich offline parser: offline-parser-9246.json is the confirmed copied-image run; its -initial capture remains separate. parser-object-layout-9246.json retains named routine bounds, disassembly, hashes and reviewed layout evidence. Comparison is explicitly against the 9644 live parser baseline, not a same-build live test. just offline-parser-check checks both builds without loading vendor code.

Offline parser pilot: offline-parser-9644.json retains the copied-image run and its comparison with the saved in-app parser fields. The -initial and -sandbox-initial files retain the blocked and first successful attempts. just offline-parser-check validates the confirmed capture and runs converter/comparison rejection tests without executing vendor code. Scope and reproduction: Offline Parser Host.

Skin XML discovery: skin-xml-relations.json retains vendor source hashes and direct parent/child locations. Regenerate with just skin-relations; validate with just skin-relations --check and python3 tools/test_skin_metadata.py from the repository root. The scanner tests malformed-file omission and quoted operators; query tests cover category overrides and CLI/MCP parity. See scope and query usage.

H4: runtime argument grammar tests links bounded historical-binary captures, frozen exact-script/stateful predictions, HTTP captures, restoration journals and interrupted runs. Inspect results with just runtime-grammar --artifact <capture> and static gaps with just runtime-parser-frontier; validate with just check-runtime-grammar. just runtime-grammar-editor joins the separately captured UI help observations and HTTP results, retaining failed control predictions. --grammar-playing tests/runtime-grammar-playing-cases.json --check describes the silent-audio fixture for asymmetric master/playback tests; it preserves loaded decks 1/2 and uses empty decks 3/4. Named action/scope fixtures and their baselines are listed by just probe-arg-forms --grammar-actions tests/runtime-grammar-action-cases.json --check and --grammar-scopes tests/runtime-grammar-scope-cases.json --check.

This folder contains small reproducible assets used to verify claims in the reference documentation.

For H4 symbols without a family mapping, start with just runtime-grammar --triage: it returns checked historical instruction anchors, dispositions and next actions without loading the full live-evidence audit. Triage does not establish behavioral coverage.

The H4 parser_effect_boolean action fixture requires Phaser already selected in deck 1 slot 1 and four unloaded, stopped decks. Inspect its guarded activation comparisons with just probe-arg-forms --grammar-actions tests/runtime-grammar-effect-boolean-confirmation-cases.json --check. The H4 report explains the historical cache-supplying callers and the consumer-specific evidence boundary.

The H4 incoming-parameter fixture reuses parser_zoom_levels. Inspect just runtime-grammar --artifact tests/runtime-grammar-incoming-pipeline-confirmation-9628.json --group incoming-zoom for the discriminating param_cast float bridge and same-source direct-chain controls. The incoming run report keeps failed and aborted captures separate from completed observations. Prefer this scoped query to reading the mutation journals; uncertain writes must not be replayed.

H4 pair-reader callers have a compact route: just runtime-grammar --callers 'IParamValuesAction::getValues(float*, float*)'. Use --artifact tests/runtime-grammar-pair-float-confirmation-9628.json --group pair-float for exact multiplication results; historical caller association does not prove the current float path, and unit-output prediction failures remain visible.

Sampler contract work: just coverage --section=Sampler --frontier gives the scoped queue; just verb sampler_volume joins the confirmed forms from sampler-contracts-9598.json. The initial and timeout captures with the same prefix preserve failed/aborted attempts and never supply contract closure. tools/test_sampler_contracts.py checks this boundary in just check.

To reproduce the stopped addressing/level fixture, run just probe-sampler-contracts --prepare "$HOME/Library/Application Support/VirtualDJ/Sampler/VDJ Contract Fixture.bank" once, then start from the original bank with no sample playing and run just probe-sampler-contracts --run --output /tmp/sampler-contracts-new.json. Preparation refuses an existing directory and generates WAV tones plus a bank XML and SHA-256 manifest; it uses no library media. Keep the owned bank installed for later passes. The runner checks its names, groups and empty slots before probing, journals every action before dispatch, never replays an uncertain action, and restores/reads back every level case and the original bank and deck selections. Bank changes are asynchronous, so restoration polls the bank name with a bounded deadline. Original bank names are redacted in persisted restoration entries.

This pass covers explicit selectors, group-versus-slot levels, unequal baselines, and deck defaults over HTTP. It does not start playback or establish pad-page, pressure/release, routing, recording or deletion behavior. Next use the same bank for the separately allowlisted playback pass below; pad-context tests must use a real pad fixture. Record results in the verb store and attach exact form/channel evidence to the existing assessment, rather than creating a parallel Markdown contract table.

just probe-sampler-playback prints the frozen cases from sampler-playback-cases.json without touching the app. With the same owned bank installed and all decks/sampler stopped, just probe-sampler-playback --run --output /tmp/sampler-playback-new.json plays quiet generated tones. It distinguishes start/continue/restart/stop using activity, elapsed strings, percentage positions, known WAV lengths and the active count. Different groups supply simultaneous target/witness samples; separate immediate count probes establish three/four-player states. Both runs use nonsense controls and explicit per-slot cleanup, never the all form under test. Levels, selections, original bank and stopped deck state are verified afterward. Existing output files are never overwritten.

The supplemental just probe-sampler-playback --default-scope prints sampler-default-playback-cases.json. Add --run --output /tmp/sampler-default-new.json to exercise unwrapped forms from stopped/playing states, with both nonsense selectors. It requires current get_deck=1, verifies default slot 9 before/after each case, and refuses to change deck selection. sampler-default-playback-9598.json closes only this observed HTTP default context; changing context, missing a state or missing controls prevents promotion. Reuse this supplement rather than rerunning the full numeric/aggregate/count suite for a default-form gap.

sampler-playback-9598.json is the focused playback join. Claims are recomputed from timed readbacks, not the recorded predictions, execute responses or a success flag. Missing controls, discrepant positions, slow snapshots and possible natural endings cannot close a form. Earlier -initial and -baseline-delay captures are retained separately: a delayed read can outlive a sample even when an HTTP request eventually succeeds. Timed queries therefore use bounded read-only waits; mutation requests are not replayed. This is HTTP transport evidence, not captured audio or proof of speaker/headphone routing.

Test assets are important evidence for the documentation project. They are kept outside the normal example folders so a test page, skin, mapper, or other repro fixture is not mistaken for a polished reference implementation.

Area Contents
ghidra-time-consumers-9246.json Read-only Ghidra export of the shared time reader and text formatter, independently checked code bounds/hashes and imported arm64 image identity. just ghidra-time-consumers; conditional formatter-pointer interpretation. Tier 2, no behavior promotion.
tail-consumers-shared-9246.json Expanded time-consumer pilot: same-action caller bindings, position-aware exact/prefix comparisons, shared code and unresolved formatter/output-argument paths. Default for just tail-consumers; original pilot retained. Scope and next boundary.
tail-consumers-9246.json, tail-probe-9246.json Named parameter/comparison provenance anchored to a fresh memory capture, plus native and HTTP nonsense controls. Structural candidates remain separate from recognition and behavior; initial attempts are retained. just tail-consumers VERB, --helpers, --leads, and just tail-consumers-test; method and limits.
lyrics-cache-9246.json, lyrics-binary-vectors.json, lyrics-cache-observation.json Audio-derived lyric keys, timestamped payloads, historical cache reader/writer/server paths, synthetic original-instruction vectors, and aggregate offline validation. Tier 2; no private lyrics committed. just lyrics-cache --evidence, just check-lyrics-cache; see Lyrics Cache.
linked-sid-9246.json, linked-sid-vectors.json, linked-sid-observation.json Linked-track SID: symbol-bounded historical ARM64 evidence and vendor Unicode tables, synthetic vectors executed in Unicorn, and privacy-preserving offline database comparison. All Tier 2; no live write/readback claim. just linked-sid --evidence, just check-linked-sid; regeneration and cleanup limits in Application Internals.
build-history-2026-09-06/ Historical installer binary captures: executable hashes, verb tables, architecture cross-checks, transitions, and bounded named skin-reader disassembly. Generated by tools/extract_build_history.py; structural leads only, no runtime results. summary.json → verb_history is the per-verb sampled presence and flag history that just verb joins in. vendor-text-diff.json (just vendor-history-diff) holds the older appendices' descriptions and shipped-skin usages the current app no longer carries, provenance-stamped.
Pads/ Pad-page XML harnesses used by the VDJScript local test tracker and related source notes.
Skins/ Skin fixtures used with local pad/controller/script tests.
Skins/clickthrough-probe/ Nineteen generated deck skins, identical apart from one attribute on one element, that took clickthrough from binary lead to confirmed behavior: two overlapping buttons each writing its own global, so the answer is read over HTTP rather than judged from a screenshot. Carries a calibration variant, nonsense-value and nonsense-attribute controls, and two boolean-path variants (yes, TRUE) sized by the named getBoolParam disassembly and confirmed live on 2026-09-07: boolean true makes the element transparent to clicks without firing its own action, a third state beside pass. A second series wraps the top button in a <panel>, a plain <group> and a <group visibility>, and stacks three layers: containers honor the attribute except a plain group, and pass-through is one layer per attribute.
todo-status/ Regression fixtures for the TODO.md status convention (tools/task_queue.py selftest, in just check). legacy-decorated.md holds the four decorated status forms the old awk selector silently skipped and every one must now error; normalized.md proves selection skips a Done entry for the first Ready one; structural.md covers a missing status, two status lines in one block, a status buried below narrative, and a duplicate task identifier. Expectations in expected.json.
skin-reader-vocabulary.json What the skin XML readers accept, as opposed to what shipped skins use or the SDK documents — the three disagree. Three anchored __text windows on one build, every string each references, classified against the other two sources. String references only: no call target is followed, so a name a reader hands to a helper is invisible. Absence is still informative — an attribute absent from the binary is a class="" template placeholder, not reader vocabulary. Written by tools/extract_skin_readers.py; query with just skin-reader <name>.
sysicon-atlas-wiki-2026-09-16.json, sysicon-atlas-consumers-9598/ Dated official atlas table and build-stamped static captures of the customicons reader, atlas loader, file selector and folder RTTI method. just sysicon-atlas joins cells to existing sysicon rendering evidence and labelled binary candidates. Consumer captures are Tier 2; internal numeric selectors do not establish exposed skin keys.
skin-classes.json Build-anchored RTTI, vtables, function-bounded factory branches and constructor stores, with per-class XML getter call sites. Tier 2 leads; child/helper reads are not an outer-element schema. just skin-classes computes totals; --questions reports Q1–Q3 findings and negatives; --unmatched classifies unmapped inventory names; --unwritten is the tracked discovery queue. Inspect --get CLASS for addresses, scopes and unvisited calls before a live canary. Generate a new capture with tools/extract_skin_classes.py --generate --output /tmp/skin-classes-new.json (numpy and capstone); --historical-app optionally adds cross-build symbol matching. Historical artifacts are not overwritten implicitly.
skin-classes-debug.json Companion capture of unique normalized cross-build function fingerprints and anchor references. Names are provisional historical matches, not current behavior. The main artifact carries structural mappings independently of these names.
skin-schema-button-9644.json Build-scoped button XML node ownership pilot. Outer/shared/child read paths, uncertainty flags, routine hashes and unresolved call frontier. Reuses the memory capture for image identity only; Tier 2, not runtime support. Query just skin-schema button; method and reproduction.
skin-node-helpers-9644.json Exact code and guarded structural models for named conditional child selection and matching-sibling replacement. tools/skin_node_helpers.py --check reproduces the evidence with capstone/numpy.
skin-schema-button-conditional-9644.json Current button node-ownership capture using the guarded helper models; the original capture stays as baseline. just skin-schema button queries this capture.
Skins/schema-condition-probe/ Synthetic conditional pos/size/up fixture, forward/reversed click readbacks, retained screenshots and verified skin/deck restoration. Previously unset probe variables remain zero. tools/skin_condition_evidence.py --check validates the recorded evidence.
verb-arg-positions.json Which argument positions a verb actually reads. Every other prober asks "is this token recognized", one token at a time, and is blind to a keyword in the second position — get_sample_info <slot> <field> errors on every single-token probe, real fields included. This holds the attested shape and varies one position within its own class; a changed answer means that position is read. reads means the value moved, not that the verb interprets the argument as documented, and ignored is a statement about this state only. Written by tools/probe_arg_positions.py.
get-time-positions.json get_time's position arguments against known distinct positions. The state the ten fixtures.py states could not provide: a cue, a loop start and a loop end that actually differ, each read by an oracle other than get_time, so 'cue1' can be separated from a token the verb has never seen. Records the controls' own behaviour — an unrecognized tail falls back to elapsed, not to the bare form — and the one conditional result, 'loopin' reading the loop start only while a loop is active. Written by tools/probe_known_positions.py.
verb-table.json Authoritative. VirtualDJ's own verb table, extracted from a serialised record array in __DATA,__data ({const char *name; uint32 id; uint32 flags;}): 1,028 records, 955 ids, 61 alias groups, 37 editor-hidden names, and every verb's Button Editor category (38 categories, from a const char *[38] name array plus a uint8 category-per-id array indexed by id+1). Membership proves a name is a verb and absence disproves it; records sharing an id are the same verb, flags==1 marking the alias spelling, flags==256 the hidden bit. Written by tools/extract_verb_table.py; query with just verb-table <name>.
action-contracts.json Structural contract per verb, read from the surviving ACTION_ class RTTI: name string → std::type_info → vtable. Gives every one of the 955 distinct verbs its implementation class (a checked 955/955 bijection), base-class family (toggle/slider/plain), context mixins (IActionPlugin, IActionSampler), vtable length, and the override matrix — execute, generic query, specialized text query — with slot meanings calibrated from HTTP-proven verbs, not hard-coded. Also carries method-body fingerprints: arg_demand_slots (E_INVALIDARG materialization — 114/116 agreement with sweep needs-args, plus 301 bare-answering query verbs flagged as taking optional args) method_strings (per-slot string refs), and keyword_candidates — 259 verbs whose methods compare an argument against literals, recovering undocumented enum arguments (`get_bpm absolute
verb-return-types.json Observed return type per query verb (Tier 1, HTTP): all 652 query verbs sampled in three read-only contexts and classified — 334 bool, 145 int, 68 float, 4 percent, 72 text, 29 untyped. Also records boolean_truth — whether the verb is true in ternary-condition position, which is independent of its value — and flags truthiness_trap for the 171 verbs that report a true-looking value while being false (get_version → 2026, false; no slider verb is ever true). Only 106/652 query verbs are boolean-true. The authority for a verb's concrete rendered type (proves can return, not always returns); 0 capability conflicts against action-contracts.json. Written by tools/sweep_return_types.py (needs live VirtualDJ); query with just verb-return-type <name>.
binary-verbs.json The structured verb list extracted from the app itself, unioning three sources: 1,012 mangled ACTION_<name> implementation classes, 812 languages.zip catalog entries, and the parser's 967-entry alphabetically sorted name table (which is where aliases live). 1,076 names, covering 998 of the 1,007 the HTTP sweep proved real; flags 63 alias_candidate names (in the table with no implementation class of their own). Membership proves a name is real; absence is NOT a disproof. Written by tools/extract_binary_verbs.py; query with just binary-verb <name>.
verb-existence-sweep.json HTTP existence probe for every verb-store name plus candidates-doc names: status (exists/unresolved), kind (query/action-only/needs-args/context-gated/false-here), the HRESULT, and a live sample value for query verbs. Written by tools/sweep_verb_existence.py; query with just verb-probe <name>. Proves existence and kind only, never behavior.
verb-arg-forms.json What a verb's tail accepts (Tier 1, HTTP). Every target verb's forms — bare, each candidate token, and ordered pairs (totals: summary.verbs × summary.forms, printed by just probe-arg-forms --check) — probed inside each named fixture, every reading measured against two nonsense controls, because an unknown argument is silently ignored so "it answered" proves nothing. A form counts as recognized only where it separates from garbage in at least one state; failing to separate is not evidence of absence, and a token recognized in any run stays recognized with not_reproduced_in noting where it did not repeat. Carries pair_shape per two-token form (first-token-wins, singles-agree, last-token-wins, beyond-singles) — only the last two are evidence of real two-token grammar. Guards: disputed_verbs (verdict flipped between runs — get_cpu, is_using), undiscerning_verbs (matched most of an arbitrary vocabulary, so drifting rather than speaking it — get_time, param_cast, record_vu, pioneer_cue). Written by tools/probe_arg_forms.py; query with just verb-arg-forms <name>.
action-catalog.json The vendor's own prose, offline. Resources/languages.zip → English.xml → <Actions> is the same text the official verbs appendix publishes at virtualdj.com — verified verbatim for auto_bpm_transition — so 816 descriptions are readable without the network. 97 verbs document parameters, 18 promise more than one argument. The only source that says what a parameter MEANS; probes can only say a token is not nonsense. Its --cross-check diffs documented parameters against the probe and corpus artifacts three ways. Written by tools/extract_action_catalog.py; query with just action-catalog --get <name>.
vdjscript-corpus.json Every VDJScript snippet Atomix wrote, deduped with provenance (summary.by_source has the per-source counts): shipped Built-In pad/skin/samplerbank/videoskin XML (usage — what the parser demonstrably accepts), snippets quoted inside catalog descriptions (documentation — what a form means), transcriptions from the official wiki, every <map action> in the factory controller mappings and every onchange callback in the device definitions decoded from controllers.dat (source factory, added 2026-09-12 — the largest vendor corpus by an order of magnitude and the only one that exercises the mixer, EQ, timecode and hardware-display families; the decoded XML is vendor copyright and stays under the gitignored vendor/controllers/<key>/, one tree per archive keyed by app build and archive revision (18.0.9598-r2241), added by just controllers-vendor, and every file is checked against controllers-manifests/<key>.json's SHA-256 before it is read, so --check fails rather than silently mining a stale or edited tree; summary.factory_archive stamps which archive was mined, and --check verifies that one, reporting and skipping when this machine holds a different build; a snippet keeps one origin per source and attribute plus an occurrences total, since eq_low alone appears in hundreds of mappings), and the statements compiled into the app binary — the statements behind its own menus and toolbars, read from the __cstring pool (sideview 'clone', font_size 'big', show_splitpanel 'info', sideview 'automix' 'blink'), 80 of them in no other source. Also the Atomix-published add-ons under examples/*/Official-Addons/ (source addon, added 2026-09-23 — controller screen skins, the 2018 default skin, the add-on pad pages and one add-on mapper; they carry the browser, sideview, playlist and karaoke verbs a bundle skin never writes, and are graded apart from builtin because a catalog download has no bundle member for check_bundle_copies.py to hold it against; each folder's README records the add-on id and the installed zip's SHA-256). Only vendor trees are read; examples/Pads/Quarantine and this repo's own skins are excluded so our fixtures cannot masquerade as vendor evidence. Binary strings pass a statement filter (verb first, script tokens after, one marker no prose has) because set, select and color are verbs and the pool also holds ffmpeg's option help. Written by tools/extract_script_corpus.py; query with just script-corpus --verb <name>.
attested-tails.json Argument tails recovered from those snippets (summary has the token, verb and novel counts; the factory mappings more than tripled the tail set on 2026-09-12). Verbs whose first argument is a name the author chose — the var/set/toggle family, skin_panel, repeat_start timers, controllervar — contribute their shape but not their names, since a variable name is not vocabulary; at most six snippets are kept per token and shape, one per source first, with totals in tail_counts and shapes[].count. Some tokens are attested only by the binary-embedded statements (`show_splitpanel effects
binary-vocabularies.json Argument vocabularies as GROUPS, recovered from the binary as structures rather than per verb. Every other argument source is per verb, but an enumeration is usually shared — one helper matches stem names for stem_color, stems_split and effect_stems; one table holds every colour name — and per-verb analysis cannot see it: 67 of the corpus's attested tails are in the binary but outside their verb's own code. Three signals, recorded separately and weakest first: a string-pool run (neighbours, never promoted alone), a code region that references two or more members (a switch walking the enumeration; every other word it references is a candidate), and a pointer table in __DATA_CONST (the serialised structure itself, entries in order). 21 groups, 265 members, 190 unknown to every per-verb source. Pointer tables found: 38-entry colour names (beige, marine, violet, transparent, reset …), the 5 stems (vocal hihat bass instru kick), sideview pages, the 19 settings pages (settings 'audio' and its siblings), the 17 pad-page names, the 17 song fields. Tier 2, discovery only: a member is a lead for probe_arg_forms.py, and verbs silently ignore unknown words, so confirmation needs a fixture where the forms would differ. Written by tools/extract_binary_vocabularies.py; query with just binary-vocab --get <group> / --verb <name> / --tails.
corpus-parse-results.json The corpus run through /query, and the first check here that can falsify a grammar claim. 1,167 of 1,516 parse and none contradicts VDJScript Grammar; the failures map the channel's boundary instead — 33 surface-gated action verbs, 15 pipeline forms (& param_* wants a value HTTP cannot supply), 14 execute-position semantics (loop 50% multiplies rather than sets; the binary's sampler_bank +1, pad_page +1 and deck master sampler_bank -1 are relative steps, which query position rejects with E_INVALIDARG), 29 E_FAIL (silence, not denial). All five binary statements written with && (browser_window sideview && sideview 'sidelist') parse, so the parser accepts && as a conjunction on this channel. Query-only by design: the corpus contains load, unload and browsed_song color. Re-run 2026-09-12 on build 18.0.9598 after the factory mappings joined the corpus: the structural class grew in proportion and is the same execute-position family — hot_cue 5, sampler_stop 14, delete_cue 4, pitch_range -1, sideview +1 are actions with a slot or a relative step, which /query rejects with E_INVALIDARG — with no new construct among them, so the grammar doc still stands unfalsified; the run's own summary.outcomes carries the counts. Re-run 2026-09-24 on build 18.0.9644 with the addon source in the corpus: all 20,610 snippets replayed, parsed 12,790 -> 14,875 and structural 277 -> 276, so the 2,563 add-on snippets introduced no new execute-position family and nothing contradicting the grammar doc. One snippet did land outside every known family and was probed before being classified: zoom 'save' (American Audio 10MXR v8) answers E_INVALIDARG, while zoom alone answers 0.18 and zoom 'recall' answers no — a write keyword in query position, the same boundary as zoom 2. The nonsense control zoom 'zzqqx' answers E_INVALIDARG as well, so the rejection does not separate save from junk and is evidence about the channel, not the keyword; the new state-write-keyword family is deliberately the one attested keyword, not all quoted tails. Written by tools/check_corpus_parses.py.
verb-execute-forms.json Tails measured in EXECUTE position, where the query path is blind — deck all broadcasts on execute and collapses to deck 1 on query. Each form is run from two baselines (off and on), which is what separates set (yes,yes) from flip (yes,no) from no-op (no,yes); from one baseline beatlock on and beatlock <junk> are indistinguishable. Writes to a live instance, so: allowlist only, each verb round-trip tested first, every value restored and verified, abort on a failed restore. Written by tools/probe_execute_forms.py.
action-vtables-9246.json Every ACTION_ vtable slot resolved to a named method, from bundle 18.0.9246 — the last unstripped build (321,571 symbols; 18.0.9482 onward keep ~1,210). Names the 12-slot layout the current build can only infer: 2 onExecute, 3 onQuery, 4 onQueryBool, 5 onQueryText, 6 onTooltip, 7 onUp, 8/9 onSaveState/onLoadState. overridden_slots agrees 950/952 with the current build, so names transfer — but a single verb's contract still comes from the current-build RTTI. Written by tools/extract_action_vtables.py.
action-modules-9246.json Which of Atomix's own source modules implements each verb, from the STABS debug symbols in the same unstripped bundle 18.0.9246: dsymutil -symtab N_OSO records name each object file, N_FUN records the functions linked from it. A complete partition — 952 ACTION_ classes over 37 action_*.cpp modules on build 18.0.9246, each class in exactly one, all 952 matching a store verb or alias (so the binary hides no verb the store lacks). Vendor build metadata about where code lives: Tier 2. It is strong evidence of how Atomix GROUPS a verb and no evidence that the verb works — nothing here may set a test status. The module is coarser than the store's section: action_audio splits into Deck Management and Audio Playback, action_param three ways, and action_get is a grab-bag of 117 queries landing in three sections. --sections grades each module — clean (its sectioned members agree ≥90% on ≥5 verbs), mixed, too-few, no-evidence — clean modules named a section for their unsectioned members first (31 records, 2026-09-06); mixed modules followed on 2026-09-14 with the module's majority section, Hidden Button Editor excluded and a get_* verb taking the section its module's own get_* siblings unanimously carry (124 records, each evidence line naming the module and its agreement). too-few and no-evidence (action_prelisten) stay unsectioned, as do verbs with no ACTION_ class on b9246. Written by tools/extract_action_modules.py; query with just action-modules --get <name> / --module <name> / --sections. Regeneration needs the b9246 pkg expanded, not the installed app.
vdjremote-opener.bin Raw opening handshake captured from a real VirtualDJ Remote device (app build 8515), the reference input for tools/vdjremote_dial.py --decode and the prefix reused by tools/vdjremote_subscribe.py. Backs docs/Remote Protocol.md.
vdjremote-actions.log Decoded device→desktop action frames captured while a real device's controls were touched, including the timed press sequence that maps play/cue/crossfader/volume to control ids.
Stems/ Stem-file generators for docs/Stem File Format.md claims. make-diagnostic-stems.zsh builds probe files where the master is white noise and each stem a distinct sine tone (100/200/400/800/1600 Hz = kick/bass/vocal/instruments/hihat), so pad→stream routing and master↔stems switching are verifiable by ear; includes a stamped vs unstamped writing-application sidecar A/B.

Pad Fixtures

File Purpose
Pads/Reference - Sparse Helper Tests.xml Sparse official helper checks such as connect, system, open_stem_creator, and dualdeckmode_decks.
Pads/Reference - Statement Branch Probe.xml Calibrate debug branches against true, false, valid text and unknown names; not a general validity test. See Statement Branch Probe.
Pads/Reference - Sampler Loaded Test.xml Page-aware sampler loaded-state checks.
Pads/Reference - Hidden Button Editor Tests.xml Low-risk flag1-hidden Button Editor taxonomy probes such as stem_volume, sampler_inputgain, pad-page helpers, and plugin query helpers.
Pads/Reference - Mix FX Query Test.xml Mix FX selected-state and active-state query checks.
Pads/Reference - Deck Error Test.xml Deck load-error state checks for deck_has_error.
Pads/Reference - Dual Deck Mode Test.xml Current and deck-scoped dualdeckmode_decks checks.
Pads/Reference - FX Introspection Test.xml Deck FX slider/button label, count, default, and effect_has_* checks.
Pads/Reference - Grammar Battery Test.xml VDJScript grammar checks: ternary/& precedence, nested-ternary associativity, and backtick-computed arguments vs param_* chaining.
Pads/Reference - FX Bank Test.xml FX bank save/load restore checks for slot names, active state, slider values, and deck scope.
Pads/Reference - Release FX Test.xml Release-FX slider and is_releasefx checks against normal slot slider behavior.
Pads/Reference - BeatGrid Command Test.xml BeatGrid-specific effect_command probes based on the built-in plugin UI XML.

Skin Fixtures

Folder Purpose
Skins/MixFxQueryTest/ Skin-side Mix FX query checks used with the Mix FX pad fixture.
Skins/PlaceholderConditionTest/ Define-placeholder visibility="" / condition="" canary for skin XML.

Bounded verb-tail leads

action-tail-leads.json is a regenerable Tier-2 probe queue joining bounded method/helper traces, existing vocabulary groups, and historical method/module annotations. Helpers carry fan-out and verb-id-fraction measurements so the script evaluator's dispatch dump ranks below verb-specific finds, and the verb store is joined live at query time so tails it already records leave the queue. Query with just action-tail-leads --get NAME or --queue; summary totals are available from just action-tail-leads → summary. See Verb Tail Structural Discovery for the transfer assessment, regeneration order and live-probe limits.

action-contracts-debug.json holds the call-graph addresses behind those traces — roots, direct edges, visited and unvisited targets, per class and slot. It is written by the same extraction and read with just verb-traces NAME; the contracts check fails if the pair disagree about the binary they came from.

Long-track time-reader capture

tests/long-time-forms.json is the task 10d HTTP capture. just long-time-forms reports verdicts; just long-time-forms get_time_hour selects a verb. The source tools/probe_long_time.py generates temporary FLAC audio with ffmpeg, verifies duration with ffprobe, and requires empty stopped deck 1. just probe-long-time changes position, pitch and display mode, then verifies restoration after each independent run. It writes only after all runs restore successfully.

just check recomputes verdicts from raw readings and verifies the hour arithmetic against position, pitch and generated duration. The view joins this artifact without relabeling legacy measurements. A documented elapsed fallback is distinguished from a token that separates from controls; unchanged sign readings remain unresolved.

Controller archive reader and SIMPLE_MIDI fixture

just controllers-vendor decodes the installed build's controllers.dat under its archive key (just controllers-archives lists them, just controllers-diff A B compares two builds from the committed manifests in controllers-manifests/); just controllers-extract is the ad-hoc decode; just controllers --device APC64 and just controllers --path /device/settings query the generated manifests without reading the entire corpus. See Compiled Controller Definitions for container layout and the live fixture for the custom-definition load test, byte sequences, GUI readback, and cleanup.

Reader regression checks: uv run --with pycryptodome==3.23.0 python tests/test_read_controllers.py. They cover the real RSA envelope, exact XML preservation, chained blocks, CRC corruption, truncation, path safety, duplicate names and malformed XML.