Merge/master 0.72.0

[markwpearce]

Brings 0.71.0 → 0.72.0 of master into v1. Master added 33+ features and fixes
(141 non-merge commits) that needed to be reconciled with v1's reworked
architecture. Tests: 4035 passing / 9 pending / 0 failing. Lint clean.

What v1 changed that affected the merge

v1's core architecture differs from master in ways that touched almost every
PR being brought across:

• pkgPath vs destPath — v1's pkgPath holds the transpiled
  extension (.brs), while master's held the source extension. v1 added
  destPath for the source-extension path. Anywhere master's code reasoned
  about user-authored paths (imports, file lookups, rename edits) had to be
  switched to destPath.
• parser.references → _cachedLookups — v1 moved per-file lookups
  (importStatements, functionStatements, classStatements, etc.) from
  parser.references to a private _cachedLookups field on BrsFile.
  Every consumer master added had to be re-pointed.
• Plugin lifecycle event names — v1 renamed onGetCodeActions →
  provideCodeActions, onFileValidate → validateFile, onScopeValidate
  → validateScope, beforeFileTranspile → prepareFile,
  afterProgramValidate → afterValidateProgram. New emit/listen sites
  from master were translated.
• Diagnostic shape — v1's BsDiagnostic uses location: { uri, range }
  instead of master's range + BsDiagnostic.file. Every diagnostic master
  emitted got a shape rewrite.
• Type system — v1 has a real type graph (BscType,
  getMemberType/getCallFuncType, ReferenceType lazy resolution).
  master's "rename keyword to identifier" tricks had to round-trip through
  the graph.
• Build pipeline — v1 dropped Program.transpile([{src, dest}], stagingDir)
  in favor of Program.build({ outDir }) with prepare → serialize → write
  plugin events. Anywhere master code or tests called the old API, they
  were redirected through the new pipeline.
• BrsFile.serialize() returns map as a JSON string, not a
  SourceMapGenerator. Master's chaining helpers operated on the generator;
  v1's equivalents work over the JSON.

Master features brought in and how they landed in v1

Diagnostic suppression (#1699, #1659, #1681)

• bs:disable / bs:disable-line / bs:disable-next-line directives plus
  bs:enable re-enable blocks.
• "Disable {code} for this line / for this file" quick fixes.
• "Source fix-all" code action (the umbrella action that fires every
  applicable fix-all in one shot).
• Per-node const-cycle diagnostic (was a single per-cycle diagnostic).

v1 work needed: the entire CodeActionsProcessor was a wholesale port
(911 lines) onto v1's _cachedLookups / destPath / location-shape
diagnostics. The findExistingDisableDirectives walk was translated onto
v1's BrsFile comment-flag model. Quick-fix titles thread through
testGetCodeActions which now filters out the universal "Disable …"
suggestions so legacy tests keep passing.

Auto-imports on file rename (#1688, #1671 follow-up)

• LSP willRenameFiles → Project.getFileRenameEdits returns
  FileRenameTextEdit[] (extends LSP TextEdit).

v1 work needed: master's helper compared oldFile.pkgPath against
the resolved import path. v1's pkgPath carries the .brs extension, but
imports reference .bs source paths, so the comparison and the
util.getPkgPathFromTarget call were switched to destPath. The private
computePkgPathForNewSrcPath helper was renamed
computeDestPathForNewSrcPath (it was already returning the source path
despite its name).

Selection ranges (#1657)

LSP selection-range provider — outermost expression, then statement, then
block, etc.

v1 work needed: added ProvideSelectionRangesEvent /
BeforeProvideSelectionRangesEvent / AfterProvideSelectionRangesEvent to
interfaces.ts so the bsc plugin's processor could emit through v1's
plugin event system instead of master's direct call.

Manifest reload (#1700)

Project reloads when its manifest's contents change.

v1 work needed: restored Project.activate reading
manifestFileContents (had been commented out during conflict resolution),
which is what the change-detection in ProjectManager.handleFileChange
compares against.

Sourcemap chaining + portable maps (#1676, #1624)

• Chain a prebuild input sourcemap (co-located .map, sourceMappingURL
  comment, or inline base64 data URI) through transpile.
• relativeSourceMaps BsConfig option to make sources[] relative to the
  output map (with optional sourceRoot).

v1 work needed: master's _chainInputSourceMap operated on a
SourceMapGenerator before serialization; v1's BrsFile.serialize() already
hands back a JSON string, so a new util.chainInputSourceMap(json, file)
parses, applies the input map via SourceMapConsumer round-trip, and
returns the chained JSON. The bsc FileSerializer calls it for both .brs
and .xml outputs. Master's serializeSourceMap ran during transpile
when the absolute output path was known; v1 only knows that path at
write-time, so the relativeSourceMaps / sourceRoot sources[] rewrite
was moved into FileWriter.

Line continuation in .brs (#1667, #1693)

• +, boolean ops, relational ops, and function-call-arg comma can span
  newlines. Always allowed in .bs, allowed in .brs when
  minFirmwareVersion >= 15.3.
• minFirmwareVersion BsConfig option (Support minFirmwareVersion in bsconfig.json #1678).

v1 work needed: v1's parser already had allowLineContinuation. Just
needed the firmware-version gate and to preserve the consumeNewlinesIfAllowed
calls inside finishCall / binary expression parsing through v1's call-site
restructure.

Cross-file const / enum inlining (#1680, #1578)

• Resolves const X = something.const_or_enum_ref recursively across files.
• Detects const cycles at validation time.

v1 work needed: master's recursive-inlining lived in
BrsFilePreTranspileProcessor; v1 had renamed that file to
BrsFileTranspileProcessor (no "Pre"). The processInlinedConstValue /
processExpressionForInlinedValue helpers were ported into v1's processor
and fire when a const value is non-literal so inner enum/const refs still
inline. Const-cycle detection threads through v1's per-node diagnostic
model.

validate BsConfig flag (#1687)

Allows skipping the validation phase entirely.

v1 work needed: wired into Program.validate and ProgramBuilder's
build pipeline alongside v1's existing noEmit / validate: false opts
(which were already set on the LSP path; this is the user-facing version).

Project activation concurrency limit (#1627)

• Cap on parallel project activations during sync.
• Removed void this.validate() from Project.activate.

v1 work needed: removing the eager validate uncovered v1's
"cancellation marks files as isValidated" bug — the cancelled
mid-flight validation left files flagged validated, so the follow-up
validation thought there was nothing to do. Fixed in Program.validate's
onCancel handler by rolling back per-file isValidated.

Per-file namespace contributions, lazy (#1684, #1683)

• ScopeNamespaceLookup is a Map<string, NamespaceContainer> view that
  builds entries on demand from a per-program contributors map.
• Optional fields on NamespaceContainer are lazy-allocated.

v1 work needed: v1 already had _allNamespaceTypeTable (per-scope
aggregate symbol table). Master's lazy-view sits above that. Added
namespaceStatements to both NamespaceContainer and
NamespaceFileContribution; BrsFile.getNamespaceContributions populates
it; Scope.linkSymbolTable uses it to wire each per-namespace-statement
SymbolTable to the aggregate's sibling chain.

mergeSymbolTable reference sharing (#1682)

Source symbols are shared by reference across the merge.

v1 work needed: v1's mergeSymbolTable had a placeholder body that
created empty arrays without populating them. Replaced with
destSymbols.push(...sourceSymbols). Two ride-along bugs surfaced:
(a) the merge didn't honor data.doNotMerge, so typecast m as Thing1
inside one Alpha.Beta block leaked through the per-namespace aggregate
into every other Alpha.Beta block — restored the doNotMerge filter
that the old commented-out version had; (b) getSymbol's
addAncestorInfo wrapper was cloning every result including own-table
hits, breaking the reference-identity contract — narrowed the clone to
only fire when memberOfAncestor is actually true.

LSP perf (#1628, #1626)

• Per-project PathCollection cache during flushDocumentChanges.
• onDidChangeWatchedFiles debouncing.

v1 work needed: straight ports — the underlying caches and queues
already existed in v1's LSP layer with the same names.

Import-statement definition provider (#1595)

Cmd-click on an import path jumps to the imported file.

v1 work needed: new event handler in BrsFileDefinitionProvider
operating on v1's _cachedLookups.importStatements (master used
parser.references.importStatements).

Smaller fixes brought across

• Typedef namespace param fix #1641 typedef-with-namespaced-default-param fix.
• Fix typdef generation of default param func #1551 typedef of default param func.
• Support transpiling class methods as named functions #1548 transpile class methods as named functions.
• Flag param names that are reserved words #1556 flag reserved-word param names.
• Fix for adding files on beforeProgramValidate #1568 beforeProgramValidate plugin can add files.
• Fix confusing diagnostic when dottedGet follows function call in ExpressionStatement #1598 better dottedGet-after-function-call-in-ExpressionStatement diagnostic.
• Fix crash when bsc plugin in worker loads another version of bsc #1579 plugin-version-mismatch crash fix.
• Fix signature help crash on malformed function declarations #1536 signature-help crash on malformed function declarations.
• Add projectDiscoveryExclude setting and files.watcherExclude support #1535 / Add max depth configuration for Roku project discovery #1533 / Rename setting to enableProjectDiscovery #1525 / Support projects array in settings #1521 / Add enableDiscovery language server option #1520 / Improve manifests discovery #1518 / Improve bsconfig.json auto-discovery #1512 — project
  discovery tuning (projectDiscoveryExclude, enableProjectDiscovery,
  max depth, projects array, manifest discovery).

Real bugs surfaced by the merge and fixed inside it

• TypeStatement transpile emitted a leading commented copy of the
  source ('type number = dynamic) instead of stripping the alias —
  matched master's behavior of returning [].
• Parser error recovery for unclosed call args — the
  unmatched-left-token diagnostic landed on \n instead of the
  surprising terminator. finishCall now swallows newlines before block
  terminators (end function, end sub, EOF) so the diagnostic falls on
  the right token and the surrounding block emits its own
  expectedNewlineOrColon.
• Cancelled validation marks files as validated — see "concurrency
  limit" above; rolled back in onCancel.
• Typecast m namespace bleed — see "mergeSymbolTable" above.
• Standalone-project test race — Project.activate's eager
  void this.validate() was racing the sync phase's awaited validate
  and cancelling itself with files half-validated; combined with the
  isValidated rollback, the "clears standalone file project diagnostics"
  test now passes.

Test parity

Every it.skip / describe.skip on the branch matches a skip that was
already on master or v1. No tests became newly-skipped during the merge.
Stripped diagnostic assertions on 8 line-continuation specs were restored
using v1's equivalent diagnostic name (expectedStatement vs master's
expectedStatementOrFunctionCallButReceivedExpression).

Notable design choices kept from v1

• Composite type fallbacks (X or Y, (X and Y)[], type alias to a
  union) transpile to dynamic rather than master's object.
• roSGNode* / ro*Component / built-in interface type names transpile
  to dynamic (and skip cannot-find-name validation when used as types).
• Sourcemap chaining and the relativeSourceMaps rewrite both operate on
  the JSON that BrsFile.serialize() produces, not on a live
  SourceMapGenerator.