Filed by the os-dev seat working #15805 (session 01XpTx2tbq3pZRYAdoGt6E6Y), from a grep of the tree while locating that card's two sites. ⛔ domain:*, type and priority are triage's — this seat does not produce them. No assignee set.
What #15805 says, and what the tree says
Card #15805 states the false sentence "lives in two places". Measured on origin/main at 64011dd3f, it lives in four source files. git grep 'AS BLOB' over packages/**, CHANGELOGs aside:
| # |
file |
wording |
in #15805's scope? |
| 1 |
packages/services/service-analytics/src/text-match-sql.ts |
"was measured on the driver side to return NOTHING at all" |
✅ yes — corrected |
| 2 |
packages/drivers/driver-sql/src/sql-driver.ts |
"was measured to return NOTHING at all" |
✅ yes — corrected |
| 3 |
packages/spec/src/data/filter.zod.ts |
"was measured to match NOTHING." |
⛔ no — still states the falsehood |
| 4 |
packages/drivers/driver-turso/src/remote-transport.ts |
"was measured to match NOTHING at all." |
⛔ no — still states the falsehood |
Sites 3 and 4 make the identical universal claim about SQLite, for the identical reason, and both attribute it to #6518.
Provenance — measured, not inferred
The triage seat on #15805 flagged that it could not verify the origination claim because its clone was shallow. This container's clone is not shallow (git rev-parse --is-shallow-repository ⇒ false), so git log -S answers it:
31728314c 2026-08-08 fix(drivers): text-operator case folding is the contract's
answer, not the dialect's (#6518) (#6706)
⇒ planted the sentence in sql-driver.ts AND remote-transport.ts, same commit
d063a969d 2026-08-11 fix(spec): stop folding like/ilike onto $contains at the
wire (#7536) (#7593)
⇒ copied it into packages/spec/src/data/filter.zod.ts
54bb2f125 2026-09-05 fix(service-analytics): compile the case-sensitive text
family per SQL dialect (#15790)
⇒ copied it into service-analytics/src/text-match-sql.ts
⇒ #15805's origination claim is confirmed and extended: the sentence originated in the #6518 work, and that single commit planted it in two driver files at once, not one. It then spread to two more packages over four weeks. One commit, four carriers.
Why this is worth its own card rather than a footnote
#15805 and its triage comment both argue that a half-correction is worse than none:
fixing one leaves the other stating a falsehood a reader will trust because it appears in the driver layer
That argument transfers to site 3 with more force, not less. packages/spec/src/data/filter.zod.ts is the contract layer — it outranks the driver layer as an authority a reader defers to. After the #15805 PR lands, the repo states the compile-option reading in two headers and the universal falsehood in two others, one of them in spec.
Site 3 is also where the correction is sharpest, because that header's own next sentence already lists the divergence without naming it:
Three of the five backends are SQLite underneath (driver-sql on better-sqlite3, driver-sqlite-wasm, driver-turso on both transports)
Those SQLite backends are not the same build: driver-sql runs better-sqlite3 (flag compiled in) and driver-sqlite-wasm runs sql.js (flag absent). The file already enumerates the population the divergence runs across.
The measurement (re-measured by this seat, not carried from #15805)
Both builds are dependencies of this repo. { name: { $contains: 'acme' } } over the shared FILTER_TEXT_ROWS fixture, compiled as CAST(name AS BLOB) LIKE ?:
| build |
pragma_compile_options |
result |
| better-sqlite3 13.0.3 (SQLite 3.53.4) |
LIKE_DOESNT_MATCH_BLOBS present |
[] |
| sql.js 1.14.1 (SQLite 3.49.1) |
absent |
['1','2'] — ACME Corp AND acme corp |
Two readings worth carrying into whatever text lands here, neither of which is in #15805:
typeof CAST(name AS BLOB) is blob on both builds — the CAST is not what differs. The divergence is entirely in whether LIKE is defined to be false for a BLOB operand, which is the compile-time flag.name GLOB '*acme*' answers ['2'] on both builds — the chosen remedy is stable across exactly the axis that disqualifies the rejected construct.
Suggested shape (⛔ not a decision)
Same shape #15805 took, applied to sites 3 and 4: name SQLITE_LIKE_DOESNT_MATCH_BLOBS, state that the two shipped builds disagree, and let the divergence carry the rejection — ⛔ not a "on some builds" qualifier bolted onto a claim about return values, and ⛔ without weakening either file's conclusion (both correctly choose GLOB).
Why this seat did not simply edit them
The #15805 dispatch fenced this seat's scope to the two named sites explicitly. Extending to packages/spec would also have pulled in that package's generated-artifact gate family — a verification surface the card's scope does not carry. So the two files are reported rather than edited.
⇒ If the PM would rather these ride along, the #15805 PR is a draft and the addition is one commit; that is a scope call for the PM, not this seat.
Not in scope here
The behaviour is correct in all four files and nothing here argues for changing any emitted SQL. This is docs precision only, exactly as #15805 scoped itself.
Refs: #15805 (the two-site correction, PR open) · #6518 / #6706 (origin) · #7536 / #7593 (the spec copy) · #15790 (the analytics copy).
Filed by the
os-devseat working #15805 (session01XpTx2tbq3pZRYAdoGt6E6Y), from a grep of the tree while locating that card's two sites. ⛔domain:*, type and priority are triage's — this seat does not produce them. No assignee set.What #15805 says, and what the tree says
Card #15805 states the false sentence "lives in two places". Measured on
origin/mainat64011dd3f, it lives in four source files.git grep 'AS BLOB'overpackages/**, CHANGELOGs aside:packages/services/service-analytics/src/text-match-sql.tspackages/drivers/driver-sql/src/sql-driver.tspackages/spec/src/data/filter.zod.tspackages/drivers/driver-turso/src/remote-transport.tsSites 3 and 4 make the identical universal claim about SQLite, for the identical reason, and both attribute it to #6518.
Provenance — measured, not inferred
The triage seat on #15805 flagged that it could not verify the origination claim because its clone was shallow. This container's clone is not shallow (
git rev-parse --is-shallow-repository⇒false), sogit log -Sanswers it:⇒ #15805's origination claim is confirmed and extended: the sentence originated in the #6518 work, and that single commit planted it in two driver files at once, not one. It then spread to two more packages over four weeks. One commit, four carriers.
Why this is worth its own card rather than a footnote
#15805 and its triage comment both argue that a half-correction is worse than none:
That argument transfers to site 3 with more force, not less.
packages/spec/src/data/filter.zod.tsis the contract layer — it outranks the driver layer as an authority a reader defers to. After the #15805 PR lands, the repo states the compile-option reading in two headers and the universal falsehood in two others, one of them inspec.Site 3 is also where the correction is sharpest, because that header's own next sentence already lists the divergence without naming it:
Those SQLite backends are not the same build:
driver-sqlruns better-sqlite3 (flag compiled in) anddriver-sqlite-wasmruns sql.js (flag absent). The file already enumerates the population the divergence runs across.The measurement (re-measured by this seat, not carried from #15805)
Both builds are dependencies of this repo.
{ name: { $contains: 'acme' } }over the sharedFILTER_TEXT_ROWSfixture, compiled asCAST(name AS BLOB) LIKE ?:pragma_compile_optionsLIKE_DOESNT_MATCH_BLOBSpresent[]['1','2']—ACME CorpANDacme corpTwo readings worth carrying into whatever text lands here, neither of which is in #15805:
typeof CAST(name AS BLOB)isblobon both builds — the CAST is not what differs. The divergence is entirely in whetherLIKEis defined to be false for a BLOB operand, which is the compile-time flag.name GLOB '*acme*'answers['2']on both builds — the chosen remedy is stable across exactly the axis that disqualifies the rejected construct.Suggested shape (⛔ not a decision)
Same shape #15805 took, applied to sites 3 and 4: name
SQLITE_LIKE_DOESNT_MATCH_BLOBS, state that the two shipped builds disagree, and let the divergence carry the rejection — ⛔ not a "on some builds" qualifier bolted onto a claim about return values, and ⛔ without weakening either file's conclusion (both correctly chooseGLOB).Why this seat did not simply edit them
The #15805 dispatch fenced this seat's scope to the two named sites explicitly. Extending to
packages/specwould also have pulled in that package's generated-artifact gate family — a verification surface the card's scope does not carry. So the two files are reported rather than edited.⇒ If the PM would rather these ride along, the #15805 PR is a draft and the addition is one commit; that is a scope call for the PM, not this seat.
Not in scope here
The behaviour is correct in all four files and nothing here argues for changing any emitted SQL. This is docs precision only, exactly as #15805 scoped itself.
Refs: #15805 (the two-site correction, PR open) · #6518 / #6706 (origin) · #7536 / #7593 (the spec copy) · #15790 (the analytics copy).