Repository navigation
v1.10.0-alpha -- sparse DFM, enum casts, lint-all 17% faster
v1.10.0-alpha -- 2026-09-06
The release where a .dfm stopped being read as a complete document. Delphi
does not stream a published property whose value equals its declared default,
so an ABSENT property is not an unknown -- most of the time it is a value that
can be read straight off the declaration. The converter had been treating absent
as unknown, which made it pessimistic about exactly the properties that are most
often at their default.
No schema change, no extractor change. Schema 21,
DRAGLINT_EXTRACTOR_VERSION 1.10.0-alpha and DRAGLINT_RESOLVER_VERSION
1.1.0-alpha are all untouched: everything below is resolved at QUERY time, so
no index re-parses because of this release.
Note for anyone reading the two version strings side by side: as of this
releaseDRAGLINT_VERSIONandDRAGLINT_EXTRACTOR_VERSIONare BOTH the
string1.10.0-alpha, and that is a coincidence, not a coupling. They are
independent constants -- the extractor one moved on its own schedule and last
changed in the 2026-08-30 extractor batch. Do not infer from the match that
bumping one bumps the other.
lint-all is 17% faster, and the flow checker 42%
FlowChecker.Check was 38.3% of a lint-all run. It is now 26.6%.
Profiling the five flow ORACLES -- the store-backed predicates the dataflow
lattices ask about ownership, parameter modes and type categories -- found
three of them with no cache at all, at a 100% miss rate. IsRecordType
recomputed ResolveTypeCategory 46,236 times in a single run of this repo's
own corpus. Memoising them within a file:
FlowChecker.Check 123.79 s -> 71.54 s (-42.2%)
lint-all TOTAL 322.88 s -> 268.68 s (-16.8%)
The largest single component was one nobody had predicted: the
definite-assignment replay phase fell 32.56 s -> 1.46 s, having been
almost entirely oracle traffic rather than replay work.
Findings are byte-identical. The change was gated on a full lint-all
output comparison across the corpus -- same findings, same order, same summary
line. The memos are scoped to one file and cleared on entry to each file's
check, so no answer can cross a file or a database.
The profiling counters ship enabled and print under DRAGLINT_PROFILE; only the
printing is gated, so the measured code is the shipped code.
A project index now tells you when it holds rows it cannot refresh
A project database can contain files that are not members of that project --
historically from pointing a folder scan at a project DB, which widened its
scope permanently. Two individually correct behaviours combined badly:
index --project walks the compile closure, so it never visits a non-member;
and eviction is deliberately bounded to the project's own roots, so a shared
database never has another project's rows silently deleted.
The result was a freshness warning that could never be cleared, on every run,
advising a command that could not possibly work. On this repo's own self-index
that was 81 of 199 rows.
index --project now reports them, without deleting anything:
scope: 81 indexed file(s) are outside this project's compile closure
(NOT evicted -- they lie outside every eviction root; pass --rebuild to drop them):
and the freshness note is honest about its own limits, naming --rebuild as the
only remedy that applies to those rows. Nothing is deleted: dropping them
changes what lint-all reports on, which is the operator's decision, and
--rebuild remains the way to make it.
Guard and tooling fixes
--helpverbs are now held againstREADME.mdanddocs\AI-USAGE.md.
The docs-in-sync rule always named all three surfaces, but nothing checked the
two prose ones against the verb list -- which is how--castlibcame to be
documented in--helpand neither doc. Two verbs (migrate-dbs,
shared-unit) were missing from AI-USAGE and are now documented.- A test that asserts the ORDER of two output streams was reading them through
two OS pipes merged in drain order, which reordered about 10% of the time.
It now captures through a single handle.
A .dfm is SPARSE
#default is a real fallback again. A source property that is absent because it
sits at its declared default now RESOLVES, and the resolved value is emitted
explicitly rather than silently dropped. Two new report kinds carry it:
default-superseded and default-resolved.
#when now matches a resolved default exactly as it matches a streamed value,
so a mapping over an enum finally fires on that enum's own default.
mapping-source-absent is correspondingly NARROWED to the only case where
nothing can honestly be said: absent, and no clause to resolve it to.
This reverses the #else gating that downstream rule books were told about on
2026-09-04. #else now fires only when the source is absent AND has no usable
default. A #when written to catch "the enum is at its default" by way of
#else will stop firing -- write the default as an explicit #when arm.
stored is honoured. Vcl.Controls.pas declares
Color ... stored IsColorStored default clWindow; with ParentColor set,
Color is omitted regardless of value, so resolving it to the declared default
would write a wrong value AND clear the inheritance.
Enum casts EXECUTE, and .castlib has a grammar
A #link carrying a : Cast suffix now translates its value through the
.castlib's enum blocks instead of copying it verbatim. The .castlib parser
moved into the engine so it can read one, and a new --castlib <file> flag on
convert-apply and convert-reemit names it. Without --castlib, a
cast-bearing #link passes the value through UNCHANGED. A value with no map
and no else reports the new enum-cast-unmapped kind.
--castlib is now in README and AI-USAGE, where it was missing
The flag shipped in --help but appeared in neither README.md nor
docs/AI-USAGE.md, so a reader of either could not discover that a cast-bearing
#link silently passes its value through without it. Both now document it.
convert-reemitremains deliberately undocumented, and this release does NOT
change that. It is classified as an internal stage of the conversion pipeline
inrun_docs_sync_guard.ps1's$UndocumentedOnPurposelist. That guard is
two-directional and refused an attempt to document it during this release --
correctly. Open question for the owner, not decided here: session 69 gave
convert-reemita user-facing--castlibflag and the converter team may now
want to run it directly to debug a rule. If so, it should be promoted to the
public surface and the exemption dropped; until someone decides that, it stays
internal.
Fixes
- Four ways default-resolution wrote WRONG values into a form, two of them
reproduced by running the engine at exit 0: owned-part recursion resolved a
child#linkagainst the PARENT's defaults, set defaults were truncated into
malformed DFM, and#removewas not honoured. overwrite-before-readnow counts a read inside a NESTED routine.run_index_all_jobs_spawns_workersasserted a coin flip: it sampled the
process table on a run lasting 235-780 ms, while one sample costs about as
much, so it got exactly one poll. It now asserts the parallel path's own
summary line, which is deterministic.