writeGlobalUserCore now line-surgery upserts only the [core] keys it owns
into the existing file content — no read-parse-merge-emit round-trip. Hand
edits in shapes the minimal parseSimpleToml dropped (arrays, single-quoted
strings, dotted/quoted keys, \uXXXX escapes, custom sections) survive
byte-for-byte. Dead emitTomlSections helper removed.
writeModuleTomls emits core's shipped defaults under top-level [core], not
nested [modules.core] — matches where writeCentralConfig writes core deltas
and where resolve_config.py consumers read core.* from.
tomlString auto-infers TOML type from --set values: true/false → bool,
integer pattern → int, decimal pattern → float, otherwise quoted string.
Explicit-string escape hatch: --set k='"true"' writes "true" as a string.
Previously every --set value was wrapped in quotes, so --set bmm.workers=4
wrote "4" and downstream numeric comparisons broke.
upsertTomlKey's inline-comment detector now tracks double-quoted string
state via findInlineCommentStart, so a # inside a basic-string value
(legal per the TOML spec, contrary to the previous defending comment)
is no longer truncated. tomlHasKey and upsertTomlKey both use a new
isSectionHeader matcher that tolerates trailing inline comments on the
header line and walks every section occurrence — fixes mis-routing on
hand-edited files with [core] # comment headers or duplicate sections.
Python test harnesses (test_resolve_config.py, test_resolve_customization_cascade.py)
now force BMAD_HOME to the synthetic isolation path unconditionally
instead of env.get(BMAD_HOME, default), which was a no-op when the
developer had BMAD_HOME set in their shell.
Adds 21 JS assertions covering the new behaviors directly: tomlString
type inference, # inside string values, inline-comment headers, and
writeGlobalUserCore round-trip preservation of hand-edited shapes.
Tests: 405 JS + 21 Python pass. Lint, markdownlint, prettier green.
9799d10123