Skip to content

Latest commit

 

History

History
672 lines (564 loc) · 39.8 KB

Runtime Memory Probe.md

File metadata and controls

672 lines (564 loc) · 39.8 KB

Runtime memory probe

A bounded first experiment in native plugin access to VirtualDJ's process memory. The instrument is a separate passthrough effect built by the existing plugin build script; it does not run VDJIntrospect's general probe lists.

Scope and commands

  • just plugin-memory-test exercises bounded reads, invalid addresses, malformed table candidates, and rejection of a non-VirtualDJ host.
  • python3 -m unittest discover -s tools -p test_plugin_memory.py checks that the collector rejects mismatched image UUIDs, builds, records and table locations.
  • just plugin-memory-build --install builds and installs VDJMemoryProbe.bundle in the macOS arm64 SoundEffect directory. It may require restarting VirtualDJ to discover a newly installed effect. Do not restart a working session without checking whether its loaded state can be interrupted.
  • Prefer just vdj-query "get_effect_title 'VDJMemoryProbe'" after the app has discovered the installed plugin. On build 18.0.9644 this loaded the plugin and captured from OnLoad without selecting an effect slot or activating it. Alternatively load VDJMemoryProbe into an inactive effect slot and restore that selection afterward. Cached title queries may not invoke OnLoad again.
  • Captures are new files named memory-<pid>-<unix-time>.json under ~/Library/Application Support/VirtualDJ/VDJIntrospect/. Existing captures are never overwritten. A same-second duplicate load can fail to create a file.
  • just plugin-memory-check '/absolute/path/to/memory-….json' --output tests/plugin-memory-<build>.json verifies a capture and saves a new artifact. The output path must not already exist. A verification mismatch must be investigated, not bypassed. A superseded capture keeps an -initial name.

The memory plugin accepts a VirtualDJ app bundle at any installation path. It requires the main bundle identifier com.atomixproductions.virtualdj, verifies that the loaded main image is that bundle's executable (resolving symlinks), and retains the arm64 and bounded Mach-O checks. This identity check is not signature verification. For an alternate or historical app, pass its executable explicitly to just plugin-memory-check CAPTURE --binary '/path/VirtualDJ.app/Contents/MacOS/VirtualDJ'. The collector still rejects UUID, build, table-address or record mismatches. This path flexibility applies only to the read-only memory probe; it does not relax the separate private-parser probes' build guards. A successful compilation does not establish live compatibility with an untested historical build.

What is read

The plugin reads its own task with mach_vm_read_overwrite, which reports a failed read rather than requiring a direct dereference of an untrusted address. Individual reads are capped at 32 MiB. It copies the main executable's Mach-O load commands, __TEXT,__cstring and __DATA,__data, then locates the existing sorted {name pointer, id, flags} table using hot_cue as its anchor. All candidate name pointers are resolved inside the copied string section. Missing, duplicate or unsorted candidates abort the capture.

It also reads the SDK callback object's vtable pointer and its five public method addresses. Only addresses inside an executable host segment are recorded, as unslid addresses; other slots receive null. The slots follow the public IVdjCallbacks8 header: SendCommand, GetInfo, GetStringInfo, DeclareParameter, GetSongBuffer. None is invoked by this probe. Reading a method address does not establish its internal implementation or parser layout.

No heap sweep, arbitrary query list, audio samples, browser records, track metadata, private function call, code patch, or debugger attachment is performed. Section copies are temporary and never written to disk. The capture retains only the structured verb rows and bounded technical metadata.

Evidence and limits

The collector checks the loaded image's Mach-O UUID against the executable on disk, checks the build and table location, compares every name/id/flags record with a fresh extraction, and records the disk binary's SHA-256. This is a reproduction of the same structure through a live channel, not independent proof of verb behaviour, and not a claim that the entire memory image equals the disk file.

Live result — 2026-09-21, build 18.0.9644, arm64

Verified capture and HTTP run record: a read-only title query loaded the plugin. The mapped image UUID, build, unslid table address and every name/id/flags record matched the fresh same-build disk extraction. This directly establishes that the plugin could read these host memory regions. It recovered no additional verb names relative to that disk extraction. Query verification.records in the capture for the measured total.

All public callback addresses landed inside the host executable. The bounded callback trace records their numeric/text entry points and one direct-call level, with exact LC_FUNCTION_STARTS bounds and code hashes. This trace is Tier 2 disk analysis rooted in live-located addresses, not observation of a private parser call executing.

On this build, both query callbacks pass the script pointer to 0x10057d5f8 with two zero arguments, test the returned pointer, call 0x10058029c, and then use separate numeric/text routines. Both later decrement a field at returned-object +8, conditionally invoking a virtual method at vtable +8. Inference: this is a shared parser/factory followed by context setup, evaluation and release; the decrement means an object may be destroyed before the public callback returns. The exact private ABI and lifetime are not established. Do not call these addresses or read a guessed object layout based on this trace.

Reproduce the structural trace with an optional isolated capstone environment:

uv run --with capstone --python .venv/bin/python3 python tools/plugin_memory_trace.py \
  tests/plugin-memory-9644.json --output /tmp/plugin-memory-callbacks.json

For this run, Apple's llvm-objdump Mach-O disassembly ignored the requested address bounds and began at the start of __text. Use the bounded helper above instead; it decodes only exact selected function intervals and caps routine size. The base environment did not have capstone, so this optional command supplies it without changing the project's dependency set.

The follow-up parser-object experiment below tests this candidate with fixed inputs and additional code/ownership guards. It does not make these addresses a supported or general-purpose API.

For agents: use the small verified JSON artifact and the commands above. Do not read raw process dumps or create a second Markdown list of verb records.

Parser objects — 2026-09-21, build 18.0.9644, arm64

A separate experimental plugin, VDJParserProbe, now calls the parser candidate on a fixed list of is_using inputs. This expands beyond the memory-only probe's no-private-calls scope. The calling convention is supported by the current public callback call sites and the historical named IAction::create(char const*, char const**, int) signature. Current consumer code supplies the parameter-vector layout; the public callback and current deleting destructor supply the release sequence. The guard manifest records the exact image, vtable, function bounds and code hashes. The historical address is not called or transplanted.

Before calling private code, the plugin checks the loaded build, architecture, UUID and SHA-256 of every guarded code interval. It accepts only the expected ACTION_is_using vtable, a reference count of one, and a bounded parameter span. It copies only tag/payload fields and input-matching text, then atomically releases its reference and calls the verified deleting destructor. It never evaluates or executes that private object. Separately, public GetInfo evaluates the same fixed script to obtain its HRESULT. Thus this is live object inspection, not a hook observing the object created inside a particular SDK call.

The live capture completed two identical rounds within one load; the run journal records context, bundle/capture hashes and before/after checks. All four decks were stopped and unloaded, and remained so. Every expected release call returned. The earlier discovery attempt returned an empty title and produced no parser capture before restart.

Observed representation and recognition, scoped to this fixture:

  • cue, quoted cue, and both nonsense controls became text parameters in the same action class. Nonsense was retained, not discarded by this parse. Public evaluation distinguished cue (S_OK) from nonsense (E_NOTIMPL).
  • Bare is_using yielded an empty parameter vector and public E_INVALIDARG.
  • The second arguments 7, 7.5, 1000ms and 50% produced distinct integer, decimal, millisecond and percentage tags. The percentage payload was float32 0.5; the millisecond payload was float32 1000. just parser-objects prints all decoded fields from the capture, avoiding another hand-maintained table.
  • No conclusion about recency, timing, or whether the extra arguments affect behaviour follows from these results. is_using remains behaviour-Untested.
  • The rounds are repeats within one run, not independent sessions. Object layout is measured for these cases on this build, not a public or portable ABI.

Reproduction:

just plugin-parser-test
just plugin-parser-build --install
just vdj-query "get_effect_title 'VDJParserProbe'"
just parser-objects

parser-<pid>-<time>.jsonl is written in the same working directory as the memory probe. The last command validates the committed capture by default; pass a new capture path to validate another run. A cached plugin may not rerun OnLoad. Do not overwrite an in-use bundle expecting it to reload. A newly installed name required a restart in this session. Normal quit left an idle process without UI/HTTP; SIGTERM and relaunch restored the app before the first parser call. The cause of that shutdown issue is unresolved; it is not evidence of parser failure.

Failure handling is bounded: an unexpected class or shared reference count aborts rather than guessing a destructor or freeing someone else's reference. Such an abort can retain one parser allocation until app exit. The happy path was observed; private API changes can still invalidate assumptions outside the guarded intervals. This tool is deliberately not a general arbitrary-script parser service.

Next useful discovery target: the consumer's keyword comparisons, joined to these typed parameters. This run shows why dumping parsed objects alone will not enumerate valid tails: real and nonsense text both survive parsing. Keep any consumer-derived list as candidates until a discriminating runtime test validates it. Use the parser instrument for argument types, units and binding questions.

Consumer keyword pass — 2026-09-21, build 18.0.9644, arm64

just is-using-keywords joins the current consumer extraction to the native captures named and hashed in the run journal. The extractor locates ACTION_is_using through current RTTI and its query vtable slot, then records exact direct literal-setup/comparison sites and helper bodies. It accepts only a contiguous ADRP/ADD/MOV/BL pattern, checks the helper's text-tag and literal-length checks, and keeps source register/callsite provenance. This is bounded Tier 2 extraction, not transitive analysis or a complete keyword schema.

The separate VDJKeywordProbe uses public GetInfo and GetStringInfo only. It has no private parser calls and never executes the scripts it queries. Its compiled case list includes every recovered literal in first position, nonsense controls, quoted/uppercase cue, and later inaudible positions paired with nonsense at the same positions.

The captures agree across rounds and repeat triggers in one stopped/unloaded session. First-argument cue, effect, equalizer, filter, load, loop, loopsize, pads, sample and stems returned native S_OK and off text. inaudible in first position matched the nonsense controls: numeric E_NOTIMPL and empty text with S_FALSE. Bare is_using returned numeric E_INVALIDARG. Quoted and uppercase cue matched lowercase cue in this fixture.

The consumer compares inaudible against later parameter values, whereas its feature comparisons use the first parameter (see recorded parameter setup and callsite registers). This is a structural modifier lead. Both later-position forms in the live suite matched their nonsense-tail controls; their behaviour is unresolved in this fixture. Do not turn a string recovered from this routine into a first-argument keyword without checking its role. No new keyword was found relative to the existing vocabulary, and behaviour remains Untested.

Commands:

just plugin-keywords-build --install
just vdj-query "get_effect_title 'VDJKeywordProbe'"
just is-using-keywords

The first title query loaded the plugin and captured without selecting an effect slot. To repeat the fixed sweep in the loaded plugin, read its button state first, then use effect_button 'VDJKeywordProbe' 1 and explicitly restore off if the starting state was off. In this run the bare button action captured but stayed on; repeating the bare action did not turn it off. The explicit off restore was independently read back as no. The journal preserves that initial restore failure and final successful restore. Multiple callbacks can produce multiple captures; never infer capture count from the number of actions. Files are keywords-<pid>-<time>-<serial>.jsonl in the shared probe working directory.

Re-extract the structural evidence without touching the canonical vocabulary:

uv run --with capstone --python .venv/bin/python3 python tools/is_using_consumer.py \
  --check tests/is-using-consumer-9644.json
python3 -m unittest discover -s tools -p test_is_using_keywords.py

For future agents, query the compact join before opening disassembly. This pass supports the method consumer comparisons → position-aware candidates → native recognition controls. For inaudible behaviour, the next fixture must actually separate audible from inaudible use while controlling timing; another idle sweep will not answer it. For broader discovery, apply the method to a verb whose consumer vocabulary is unresolved rather than repeating this settled list.

Shared time reader follow-up — build 18.0.9644 arm64

time-sign-consumer-9644.json follows the current ACTION_get_time_sign query slot to its single direct callee. The wrapper at 0x1004f2a9c calls the shared reader at 0x1004f268c; the artifact preserves bounded instructions, literal callsites and helper hashes. This is Tier 2 structural evidence, not an argument or behaviour promotion.

The shared reader contains literal comparison sites for elapsed, remain, total, loopin, loopout, absolute, cue and to_lyrics. Parameter zero is fetched into x23; x24 points to its string storage. A later absolute comparison uses parameter one (x22). The cue helper requires at least three characters and compares the first three; the caller then passes the suffix to another routine. Thus cue is a prefix-family lead, not evidence that only the standalone token is accepted. Suffix grammar remains unresolved.

Two structural details change the probe plan:

  • Deck-data checks branch to failure before parameters are read. An unloaded deck therefore cannot discriminate this consumer's keywords, even with native HRESULTs.
  • Unmatched first-position text reaches the same continuation as elapsed. A positive elapsed-time sign is consequently a poor nonsense control. The wrapper also has a zero-output branch, which remains a live-test lead.

The read-only empty-deck capture, on live build 9644, returned HTTP error:1 for every form listed in the capture, including both nonsense controls, in both rounds. Decks 1–4 were stopped and unloaded before and after. This establishes only the measured unavailable result in that fixture; it does not prove any tail invalid, nor recover the native HRESULT from HTTP.

Reproduce the bounded extraction and empty control:

uv run --with capstone --python .venv/bin/python3 python tools/time_sign_consumer.py \
  --check tests/time-sign-consumer-9644.json
python tools/probe_time_sign_empty.py --output /tmp/time-sign-empty-new.json

The next discriminating fixture should use generated audio, remain stopped, and read zero and negative elapsed positions against positive remaining/total time, with two nonsense controls. Establish that the requested position actually took effect before interpreting sign results. Loop, cue-prefix and lyric-relative leads need separately prepared markers/content. Do not repeat the earlier positive-time sweep or promote static branches to supported syntax.

For future agents, the tail queue's full JSON includes extensive helper evidence. Project only verb names, open tails and candidate fields when choosing work; open a selected record afterwards. A full queue dump consumes context without improving target selection.

Stopped position fixture — 2026-09-22, build 18.0.9644 arm64

time-sign-positions-9644.json records two independently loaded runs using generated, silent 60-second WAV audio. Each tests stopped positions 0 ms and 1000 ms under elapsed, remain and total display modes, with forward/reversed query order and reversed phase order in the second run. get_position & param_multiply 60000 independently verifies each position. The companion journal records write intent before each action; neither actions nor uncertain writes are retried.

At position zero, both runs produced:

Form Elapsed display Remaining display Total display
bare 0 1 1
elapsed 0 0 0
remain 1 1 1
total 1 1 1
absolute 0 1 1
either nonsense control 0 0 0

At 1000 ms all tested forms returned 1. Zero is an observed return value, despite the vendor description naming only -1 and +1. remain and total separate from nonsense at zero; their measured sign is positive. Bare form follows the selected display mode. absolute follows that mode in this fixture, but this does not establish its pitch-scaling semantics. Explicit elapsed returns zero even when the display mode is remaining/total, but matches both nonsense controls; its fallback-equivalent recognition remains unresolved.

The initial aborted attempt requested goto -1000ms after establishing zero. The independent position readback stayed zero, so the runner aborted and restored the starting state. This does not establish whether the seek was clamped or the position reader hides negative positions. Negative-sign behavior remains untested. The completed runs omit that unavailable phase. All attempts verified restoration of empty/stopped state, pitch and display mode; no playback was started.

python tools/probe_time_sign_positions.py --check tests/time-sign-positions-9644.json
python -m unittest discover -s tools -p test_time_sign_positions.py
# New capture, requiring an empty stopped deck 1:
python tools/probe_time_sign_positions.py --output /tmp/time-sign-new.json

--include-negative reproduces the attempted negative setup and aborts if its position cannot be established. Future work should first resolve negative-position readback/seek behavior, rather than repeat this now-discriminating zero fixture.

Loaded long_time boundaries — 2026-09-22, build 18.0.9644 arm64

Local test over HTTP: time-sign-loaded-9644.json reuses probe_long_time.audio() and the named long_time fixture. Independent unload/reload runs in the same application session reverse phase/form order; --repeat 3 samples each form in each stopped state under each display mode. The capture embeds exact queries, readbacks, write intent and restoration.

Negative elapsed is reachable. Both goto -1000ms from zero and goto -0.013333333333% establish elapsed -1000 ms, remaining 7501000 ms. get_position and its scaled readback still return zero before start. Thus the older position-only rejection above did not establish that seeking had failed. The retained runner verifies signed get_time 'elapsed' 'absolute' and get_time 'remain' 'absolute' together, retaining the position readback too. Start (0 ms) and end (7500000 ms, remaining zero) were also verified in both runs. No requested state was skipped in the confirmed capture; playback stayed stopped.

Explicit form Start Before start (-1000 ms) End Argument conclusion
elapsed 0 -1 1 UNDISCRIMINATED; matches both controls
remain 1 1 0 Recognized; separates at every boundary in both runs
total 1 1 1 Recognized; separates at start and before start in both runs
elapsed absolute 0 -1 1 UNDISCRIMINATED; matches first- and second-slot controls
Either first-slot nonsense token 0 -1 1 Elapsed fallback observed

Bare form follows display_time; explicit forms above have the same results under elapsed, remain and total display modes. absolute here is the second argument, not a one-token string containing a space, and not a standalone first-argument claim. Its recognition and pitch-scaling semantics remain open; zero pitch and sign alone do not discriminate them. No tail is marked Fail.

The initial aborted attempt and its journal retain an HTTP read timeout and verified restoration. The subsequent position-only initial capture skipped negative states despite recording signed elapsed -1000 ms; it is superseded by the confirmed signed-readback capture. Read-only queries may retry once after a timeout; writes never retry. Every attempt restored and verified the original empty/stopped deck, pitch and remaining-time display mode.

python tools/probe_time_sign_loaded.py --check tests/time-sign-loaded-9644.json
# Requires an empty, stopped deck 1; writes a new capture only:
python tools/probe_time_sign_loaded.py --output /tmp/time-sign-loaded-new.json --repeat 3

For future agents: inspect just verb get_time_sign and this compact matrix, not the embedded request journal. The focused coverage query now reads this capture. Do not repeat positive-midpoint sweeps or use get_position alone to reject negative elapsed. Next: prepare loopin/loopout markers, cue-prefix suffix fixtures, and lyric content for to_lyrics separately; these were out of scope.

Relocated historical host — 2026-09-22, build 18.0.9246, arm64

The updated memory probe produced a verified capture from the unpacked historical app. Its image UUID, build, table address and all name/id/flags records matched that executable. The HTTP observation record retains the responding plugin title and the limits of the session-state checks. This confirms memory-capture compatibility with this relocated build; it does not validate private parser calls or skin XML behaviour. Future historical captures should use --binary with the exact app executable instead of the collector's current-installation default.

Named tail-consumer pilot — 2026-09-22, build 18.0.9246, arm64

Result: memory anchoring plus named consumer dataflow is a useful way to assess tail candidates. It is not a universal valid-argument oracle. The fresh memory capture matched the historical executable's UUID, build, verb-table address and every record. The run journal brackets each plugin title query with a directory inventory, establishing that each produced a new capture in this run. The installed application was build 9644 while HTTP and the captured image were 9246: use the actual executable passed to the verifier, not the default app path.

The consumer artifact records exact routine bounds/hashes, named comparison call sites, possible parameter indices, unresolved receivers, and a direct-call inventory of parameter helpers. It follows both control-flow branches and traces IAction::getParam(int) results into parameter text or SActionParam::isTxt comparisons. Parameter indices are zero-based. This is Tier 2 disk analysis anchored to live memory identity, not a hook observing those instructions executing. The parameter accessor and string-member models remain structural ABI assumptions; no private routine is called.

The useful calibration cases are:

Consumer Structural result on build 9246 Live result in this run
filter_label name and clean compare with parameter 0. Skin, BASS, and FILTER go to CMessageEngine::getMessage. The lowercase filter comparisons have unresolved receivers. clean returned OFF; bare, name, both nonsense controls and the other tested strings returned DELAY, agreeing in both rounds. Only clean separated in this fixture.
is_using Feature comparisons trace to parameter 0; inaudible comparisons trace to parameters 1 and 2. Another filter comparison has an unresolved receiver. The fixed native suite distinguishes first-position feature tokens from nonsense. First-position inaudible matches nonsense; the later-position modifier pairs remain indistinguishable in this fixture.
get_song_event current/next comparisons trace to parameter 0; hasbeats, volume, volume_end, and remaining to parameter 1. getEvent() also examines parameter 0. Not re-probed: this pass recovers input provenance, not the full branching/fallback grammar or song-event behavior.
get_time_sign Its query delegates to named ACTION_get_time::getTime(long long&, SActionParam*&). Not re-probed: the shared consumer is explicitly retained as an unexpanded edge. An empty local literal list does not imply no tails.

The native capture uses the existing VDJKeywordProbe compiled suite. On this build, feature-token queries returned numeric S_OK and text off; nonsense returned numeric E_NOTIMPL and empty text with S_FALSE; the bare query returned numeric E_INVALIDARG. This supports recognition in the measured fixture, not activity/timing behavior. Both rounds agree within one capture; they are not independent application sessions. The preceding initial attempt completed both plugin captures but stopped before retaining HTTP label results because the runner's output allowlist omitted the built-in DELAY/OFF readings. Its captures remain under initial names. The corrected run completed.

All four decks were stopped and unloaded before and after. No execute calls, effect selections, restart, private parser calls or heap scan were used. Hardware/account context was not inventoried. The observation does not establish that every other part of application state was unchanged. No verb-store status was promoted from these structural or recognition findings.

Other useful targets

  • Exact public callback names. The live callback addresses resolve to CPlugin::SendCommand, GetInfo, GetStringInfo, DeclareParameter and GetSongBuffer; callback_symbols retains the same-build mapping. This makes future wrapper/context investigations easier without cross-build addresses.
  • Typed and evaluated arguments. The helper inventory exposes direct ACTION callers of getBoolParam, getFloatParam, getParamEval, getFloatParamEval, SActionParam::toFloat, toString, toColor, and the text-comparison templates. These are useful routes for deciding which consumers need unit, expression or coercion tests. Caller-site totals are queried with --helpers, not copied into prose. Shared-base calls, indirect calls and inlining remain outside it.
  • A tempting but unproven list API. getListParam(int, int&, char const*, int) is present, but the named ACTION direct-BL inventory finds no callers. Do not assume it exposes a per-verb enumeration service. Inspect shared callers and its list representation before considering a bounded private-call experiment.
  • Parameter serialization. Named SActionParam::serialize and unserialize offer leads for typed round-trip fixtures. Names alone establish neither the serialized format nor that it is safe to apply to arbitrary live objects.
  • Editor assistance and scope. DLGActionWizard::updateList, updateHint, getCurrentWord, its deckArguments data symbol, and IAction::setSource offer targeted routes for editor suggestions and source/deck binding. These remain symbol leads; the current pilot does not claim to read their data or establish their behavior.

Reuse and next experiment

just tail-consumers filter_label
just tail-consumers is_using
just tail-consumers get_song_event
just tail-consumers get_time_sign
just tail-consumers --helpers
just tail-consumers --leads
just tail-consumers-test

Queries use only the standard library. Re-extract with tools/tail_consumers.py --binary /path/to/9246/VirtualDJ.app/Contents/MacOS/VirtualDJ --check under the optional capstone environment used above. The default memory anchor is the fresh tail-probe-9246-memory.json. To repeat the read-only live calibration, use tools/probe_tail_consumers.py --binary /path/to/9246/VirtualDJ.app/Contents/MacOS/VirtualDJ --output tests/tail-probe-9246-new.json; it refuses existing outputs and requires stopped/unloaded decks. A cached or absent plugin that produces no new capture aborts rather than silently reusing old evidence.

The default report is a short human-readable view; --format=json retains callsite, receiver, routine and uncertainty details for focused follow-up.

The recommended next step is to extend this position-aware analysis through shared consumers, starting with ACTION_get_time::getTime, and carry every unexpanded call forward. Use the resulting input comparisons to select runtime fixtures. Keep localized output strings and unresolved internal comparisons out of any confirmed argument list, but do not globally disprove a token from this bounded pass. A runtime branch trace could resolve default-versus-recognized paths that return the same result, but would require a separate guarded instrument; the present memory plugin does not provide it.

Shared time consumer — 2026-09-23, build 18.0.9246, arm64

The expanded artifact now follows ACTION_get_time::getTime(long long&, SActionParam*&). The original tail-consumers-9246.json remains the initial pilot. HTTP still reported build 9246 at the start of this continuation; extraction re-verifies the existing memory anchor against the matching disk executable. No new memory capture or behavior probe was needed for this structural extension.

The named callers are get_time's numeric and text queries and the query methods for get_time_sign, get_time_hour, get_time_min, get_time_sec, get_time_ms and get_time_msf. For each selected callsite, the dataflow check requires the action receiver to survive into the shared call; an unresolved binding aborts extraction. Reports include the shared comparisons only through those verified structural bindings. They do not copy the time reader's candidates into unrelated verb reports or claim that every wrapper exposes identical semantics.

  • Parameter 0 reaches exact comparisons for elapsed, remain, total, loopin, loopout, absolute and to_lyrics.
  • Parameter 0 also reaches the named isLeftCIL prefix comparison with cue. The report marks this as a prefix family rather than an exact token. The subsequent suffix conversion and SDBInfo::getCue(int) call remain unexpanded; accepted suffix grammar and cue fallback behavior are not established here.
  • Parameter 1 reaches another exact comparison with absolute. Finding the same word in separate positions is retained as separate evidence, not merged into a flat vocabulary.
  • get_time's text wrapper compares short against a receiver whose provenance the current analysis cannot recover. The shared reader returns a parameter pointer through an output argument; stack/output-argument tracking is not modeled. The report therefore leaves short's position unresolved instead of guessing a fixed argument index. Formatting strings are classified separately.

This is Tier 2: same-build named code, bounded routine hashes and stored shared instructions, not observed runtime branch execution. Comparisons do not prove branch feasibility, keyword recognition or output behavior. Existing live time results above remain scoped to build 9644 and were not transplanted to 9246.

just tail-consumers get_time_sign now shows the shared argument candidates; just tail-consumers get_time additionally shows the text-wrapper frontier. The other named time wrappers are accepted by the same command. The default capture is now tail-consumers-shared-9246.json; --capture can still read the original pilot. just tail-consumers-test checks shared bindings, parameter positions, prefix-versus-exact classification, unrelated-verb isolation, and the deliberately unresolved formatter receiver. Re-extraction with --binary ... --check reproduces the expanded artifact.

Next: model the shared reader's output parameter and its conditional selection before assigning a position to short. Then choose a loaded synthetic fixture that distinguishes formatting and pitch-scaled versus absolute time. Repeating an unloaded time-query sweep would not settle either question.

Ghidra output-pointer follow-up — 2026-09-27, build 18.0.9246, arm64

Persistent project location

The analyzed project is now available under /Users/nom/src/virtualdj-api-reference-resources/ghidra/:

  • README.md: cache inventory and invocation examples.
  • vdj_proj_9246/: project directory; project name vdj_proj.
  • VirtualDJ-9246.arm64: imported program name and retained binary slice.
  • evidence/: raw xrefs, decompilations, symbol searches and run logs.

The retained slice's SHA-256 was rechecked against the time-consumer export on 2026-09-27 and matches. Prefer this persistent location over the earlier temporary scratch project. For inspection/export, pass -readOnly -noanalysis; do not repeat import or auto-analysis. The project/database and raw cache are local, not committed evidence by themselves.

Tracking clarification, checked 2026-09-27: tools/ghidra/VDJTailExport.java is tracked (commit bc34998). The three DrawDeck/symbol/decompile scripts were present but untracked when this route was added. The cache README's claim that the tail exporter is also untracked is stale; there is no blanket Git ignore rule for Ghidra scripts in this checkout. Check git ls-files tools/ghidra rather than inferring tracking policy from a neighboring script.

Verified export and interpretation

The completed Ghidra 12.1.3 project was reused with -readOnly -noanalysis. The export contains only the shared reader and ACTION_get_time::onQueryText, including decompiled C, function bounds, guard-byte hashes, language and imported-program hash. The new bundle at /Users/nom/src/virtualdj-api-reference-resources/VirtualDJ-9246.app matches the existing universal-binary hash and memory-anchor UUID. Its independently extracted arm64 slice matches Ghidra's imported-program SHA-256. Both guarded code intervals match the earlier shared-consumer artifact byte for byte. The target decompilations completed without target warnings; this is not a claim that the project's entire auto-analysis was error-free or that inferred C types are authoritative.

Structural interpretation, conditional on a successful shared read: the returned SActionParam* is selected as follows (positions below are one-based):

Input condition Parameter pointer returned to the text formatter
First argument has a non-text tag First argument
First argument has a text tag; second argument does not match text absolute Second argument
First argument has a text tag; second argument matches text absolute Third argument

The assembly retained in the earlier artifact cross-checks the decisive route: x22 initially receives parameter 1; the non-text path sets it to parameter 0; on the text path, csel x22, x23, x22, ne at 0x1004d6a7c selects parameter 2 when the second-position comparison succeeds. str x22, [x20] at 0x1004d6aa8 writes the selected pointer through the output argument. These are zero-based parameter indices, unlike the table's human-readable positions.

The text wrapper uses this returned pointer for its short comparison. The comparison is reached only when the computed hour component is zero; it selects the minutes/seconds format without tenths. With a nonzero hour component, the wrapper selects the hours/minutes/seconds format without checking short. It also rejects integer/value-tagged returned parameters before formatting. This explains why a single fixed argument position for short would be misleading.

The next test candidates, not promoted supported forms, are get_time elapsed short, get_time elapsed absolute short, and get_time absolute short, paired with nonsense replacements in the same positions. get_time short and get_time elapsed zzunknowna short are useful wrong-position controls suggested by the pointer-selection route. Use a synthetic sub-hour fixture with a nonzero fractional second, then an over-hour fixture: a whole-second or over-hour-only observation can hide the relevant formatting distinction.

Additional structural leads from this export:

  • The cue prefix path passes the suffix to Ghidra-resolved _atoi, then to SDBInfo::getCue(int). This suggests testing malformed, zero and missing cue suffixes separately; it is not a validated cue grammar.
  • Either the first-position absolute path or a matched second-position absolute bypasses the final division. The other route clamps a deck float field to at least 0.5 before dividing. Identifying that field's exact scaling semantics still requires a discriminating runtime fixture.

These remain Tier 2 findings from named code. The ordinary HTTP endpoint was observed on build 9644 during this continuation, so no historical-build behavior was inferred from it and no loaded media or settings were changed. The existing conservative dataflow report still labels short unresolved: this Ghidra finding is a separately reviewable interpretation, not a newly implemented general output-pointer model.

just ghidra-time-consumers
just ghidra-time-consumers --show formatter
just ghidra-time-consumers --show reader
just ghidra-time-consumers --binary /path/to/VirtualDJ-9246.app/Contents/MacOS/VirtualDJ
just ghidra-time-consumers-test

To export from an existing analyzed project, run Ghidra's analyzeHeadless with PROJECT_DIRECTORY PROJECT_NAME -process PROGRAM_NAME -readOnly -noanalysis -scriptPath /path/to/repo/tools/ghidra -postScript VDJTailExport.java OUTPUT.json 1004d63b4 540 1004d65d0 1296. The exporter refuses an existing output, caps each guard interval and aborts on decompilation failure. The lengths are frozen guards for this build, checked against the independent artifact; do not reuse them on another image. Java avoids the PyGhidra startup dependency encountered by the earlier project scripts. Validation verifies identity/bounds, not the semantics of arbitrary edits to decompiled text.