fix(store): stop legacy path migration from evicting live files (#717)

[danmackinlay]

Fixes #717 — the parts of it that are still reproducible on main (v2.6.3).

Index-time slugging is already gone (070147d stores literal filesystem paths), but two lossy-path bugs survived it, and both still cost snake_case corpora documents today.

1. The legacy-path migration evicts live files

findOrMigrateLegacyDocument() reconstructs the path a pre-2.6 index would have stored (handelize("2026_06_16.md") → "2026-06-16.md") and adopts the row it finds. That inverse is lossy — a b.md, a_b.md and a-b.md all handalize to a-b.md — and the lookup never checked whether the row it matched belongs to a file that still exists. So it renames a live file's row onto the new document, and the live file disappears from the index.

On main today:

mkdir -p /tmp/qmd-717 && cd /tmp/qmd-717
echo hyphen     > 2026-06-16.md
echo underscore > 2026_06_16.md
qmd collection add /tmp/qmd-717 --name t717
qmd ls t717
# qmd://t717/2026_06_16.md      <- only one row; 2026-06-16.md is gone

The incremental form is worse, because it removes a document that was already indexed and searchable: index 2026-06-16.md alone, then add 2026_06_16.md and run qmd update — the hyphenated file is dropped. This is the OP's "Only one file is shown", and @manpreetmaso's Logseq journals (2026_06_16.md) are the worst case since every filename in that corpus has the collision.

Both legacy lookups now skip a candidate row whose path is still owned by a file in the current scan, so only genuinely stale rows are adopted. Legitimate migration of a pre-2.6 index is unaffected (covered by the existing migration test, plus a new one where the legacy slug collides with a live file).

handelize() is deliberately left alone. It is no longer a display or index function — its only caller is this migration, where its job is to reproduce the old path format byte for byte. Making it preserve _ would mean pre-2.6 snake_case indexes stop matching and stop migrating in place — precisely the users this issue is about. Added a doc comment saying so, since "why does this still slugify underscores?" is an obvious next question.

2. _ in a path is a SQL LIKE wildcard

Same report, different mechanism, and it returns the wrong document rather than none:

qmd get 2026_06_16.md    # printed the contents of 2026-06-16.md
qmd ls notes/2026_06     # listed notes/2026-06-… too

qmd get, qmd multi-get and qmd ls <prefix> interpolate the user's path straight into a LIKE pattern, where _ matches any character and % matches anything. Added escapeLikePattern() + an ESCAPE clause at the four call sites. I found this because the regression test for (1) failed on it.

Tests

test/path-fidelity.test.ts gains a "separator collisions (#717)" block: fresh index keeps all four of 2026-06-16.md / 2026_06_16.md / my file.md / my-file.md; adding an underscored twin to an indexed collection does not evict the hyphenated file (and each path still resolves to its own content); _ is matched literally by multi-get and ls; and a legacy row still migrates when a live file shares its slug. Plus a unit test for escapeLikePattern().

Full suite passes (952 passed / 7 skipped). Two pre-existing failures on my machine are unrelated to this change and fail identically on a clean main: a batch of cli.test.ts cases that assert stderr === "" and pick up Node's DEP0205 warning under this Node version, and mcp.test.ts > MCP HTTP Transport (SqliteError: no such column: T.name). Both go away with NODE_OPTIONS=--no-deprecation / are environmental.

Not addressed here

--full-path still falls back to the qmd:// URI when a result can't be resolved on disk, and drops the docid while doing it (get/multi-get keep it). With literal paths stored, that fallback now only fires for files that genuinely no longer exist, so it reads as a stale-index signal rather than a normalization bug — happy to follow up separately if you want it to warn or keep the docid.

[PowderAddicts]

Dear tobi/qmd, We would like to acknowledge that we have received your request and a ticket has been created. A support representative will be reviewing your request and will send you a personal response.(usually within 24-48 hours). Thank you for your patience.

…

On Fri, Jul 24 2026, at 08:55 PM, tobi/qmd ***@***.***> wrote: Fixes #717 — the parts of it that are still reproducible on (#717) main (v2.6.3). Index-time slugging is already gone ( stores literal filesystem paths), but two lossy-path bugs survived it, and both still cost snake_case corpora documents today. (070147d) 070147d 1. The legacy-path migration evicts live files findOrMigrateLegacyDocument() reconstructs the path a pre-2.6 index would have stored ( handelize("2026_06_16.md") → "2026-06-16.md") and adopts the row it finds. That inverse is lossy — a b.md, a_b.md and a-b.md all handalize to a-b.md — and the lookup never checked whether the row it matched belongs to a file that still exists. So it renames a live file's row onto the new document, and the live file disappears from the index. On main today: mkdir -p /tmp/qmd-717 && cd /tmp/qmd-717 echo hyphen > 2026-06-16.md echo underscore > 2026_06_16.md qmd collection add /tmp/qmd-717 --name t717 qmd ls t717 # qmd://t717/2026_06_16.md <- only one row; 2026-06-16.md is gone The incremental form is worse, because it removes a document that was already indexed and searchable: index 2026-06-16.md alone, then add 2026_06_16.md and run qmd update — the hyphenated file is dropped. This is the OP's "Only one file is shown", and @manpreetmaso's Logseq journals ( (https://github.com/manpreetmaso) 2026_06_16.md) are the worst case since every filename in that corpus has the collision shape. Both legacy lookups now skip a candidate row whose path is still owned by a file in the current scan, so only genuinely stale rows are adopted. Legitimate migration of a pre-2.6 index is unaffected (covered by the existing migration test, plus a new one where the legacy slug collides with a live file). It is no longer a display or index function — its only caller is this migration, where its job is to reproduce the old path format byte for byte. Making it preserve handelize() is deliberately left alone. _ would mean pre-2.6 snake_case indexes stop matching and stop migrating in place — precisely the users this issue is about. Added a doc comment saying so, since "why does this still slugify underscores?" is an obvious next question. 2. _ in a path is a SQL LIKE wildcard Same report, different mechanism, and it returns the wrong document rather than none: qmd get 2026_06_16.md qmd ls notes/2026_06 # printed the contents of 2026-06-16.md # listed notes/2026-06-… too qmd get, qmd multi-get and qmd ls <prefix> interpolate the user's path straight into a LIKE pattern, where _ matches any character and % matches anything. Added escapeLikePattern() + an ESCAPE clause at the four call sites. I found this because the regression test for (1) failed on it. Tests test/path-fidelity.test.ts gains a "separator collisions ( #717)" block: fresh index keeps all four of (#717) 2026-06-16.md / 2026_06_16.md / my file.md / my-file.md; adding an underscored twin to an indexed collection does not evict the hyphenated file (and each path still resolves to its own content); _ is matched literally by multi-get and ls; and a legacy row still migrates when a live file shares its slug. Plus a unit test for escapeLikePattern(). Full suite passes (952 passed / 7 skipped). Two pre-existing failures on my machine are unrelated to this change and fail identically on a clean main: a batch of cli.test.ts cases that assert stderr === "" and pick up Node's DEP0205 warning under this Node version, and mcp.test.ts > MCP HTTP Transport ( SqliteError: no such column: T.name). Both go away with NODE_OPTIONS=--no-deprecation / are environmental. Not addressed here --full-path still falls back to the qmd:// URI when a result can't be resolved on disk, and drops the docid while doing it ( get/ multi-get keep it). With literal paths stored, that fallback now only fires for files that genuinely no longer exist, so it reads as a stale-index signal rather than a normalization bug — happy to follow up separately if you want it to warn or keep the docid. You can view, comment on, or merge this pull request online at: #784 (#784) Commit Summary a73f4b9 fix(store): stop legacy path migration from evicting live files (#717) (a73f4b9) File Changes ( 5 files) (https://github.com/tobi/qmd/pull/784/files) M CHANGELOG.md (19) (https://github.com/tobi/qmd/pull/784/files#diff-06572a96a58dc510037d5efa622f9bec8519bc1beab13c9f251e97e657a9d4ed) M src/cli/qmd.ts (15) (https://github.com/tobi/qmd/pull/784/files#diff-e7252cdb14bd10e7feaedc038d3506947354fe9b644cd034e37a7f39c580859f) M src/store.ts (68) (https://github.com/tobi/qmd/pull/784/files#diff-2717c7fb31c2070dd96f07394997809bc9be7f3fff03ab638db696cc480f39b6) M test/path-fidelity.test.ts (155) (https://github.com/tobi/qmd/pull/784/files#diff-b64a9cad6d330085e7e182984b7a40d03d35cc68b02260c5b3c1cfa3b2f5a324) M test/store.helpers.unit.test.ts (10) (https://github.com/tobi/qmd/pull/784/files#diff-ee2e697b3b8fd5a3ffc8512c1b66a3a95181f458fcca811a807feee58e058046) Patch Links: https://github.com/tobi/qmd/pull/784.patch (https://github.com/tobi/qmd/pull/784.patch) https://github.com/tobi/qmd/pull/784.diff (https://github.com/tobi/qmd/pull/784.diff) — Reply to this email directly, view it on GitHub, or (#784?email_source=notifications&email_token=BRF3HG6JOEME34NMFQFCONT5GPENTA5CNFSNUABEM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UF42DCMRZGIZTAMRZGSTHEZLBONXW5KTTOVRHGY3SNFRGKZFFMV3GK3TUVRTG633UMVZF6Y3MNFRWW) unsubscribe. (https://github.com/notifications/unsubscribe-auth/BRF3HG6327KTBV55PCL2CBL5GPENTAVCNFSNUABGKJSXA33TNF2G64TZHMYTCMJSGM3DKMZQGE5US43TOVSTWNBZG4YTSNZSGUZTRILWAI) Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS and (https://github.com/notifications/mobile/ios/BRF3HG77PLJE3TMPNYKIGJL5GPENTA5CNFSNUABEM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UF42DCMRZGIZTAMRZGSTHEZLBONXW5KTTOVRHGY3SNFRGKZFFMV3GK3TUVJTG633UMVZF62LPOM) Android. Download it today! (https://github.com/notifications/mobile/android/BRF3HG3GIUX4Y56T2PHQPLL5GPENTA5CNFSNUABEM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UF42DCMRZGIZTAMRZGSTHEZLBONXW5KTTOVRHGY3SNFRGKZFFMV3GK3TUVZTG633UMVZF6YLOMRZG62LE) You are receiving this because you are subscribed to this thread. Message ID: <tobi/qmd/pull/784 @ github . com> [ { ***@***.***": "http://schema.org", ***@***.***": "EmailMessage", "potentialAction": { ***@***.***": "ViewAction", "target": "#784?email_source=notifications\u0026email_token=BRF3HG3R22OFDOCG5WUIQ5D5GPENTA5CNFSNUABEM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UF42DCMRZGIZTAMRZGSTHEZLBONXW5KTTOVRHGY3SNFRGKZFFMV3GK3TUVNTW2YLJNRPWG3DJMNVQ", "url": "#784?email_source=notifications\u0026email_token=BRF3HG3R22OFDOCG5WUIQ5D5GPENTA5CNFSNUABEM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UF42DCMRZGIZTAMRZGSTHEZLBONXW5KTTOVRHGY3SNFRGKZFFMV3GK3TUVNTW2YLJNRPWG3DJMNVQ", "name": "View Pull Request" }, "description": "View this Pull Request on GitHub", "publisher": { ***@***.***": "Organization", "name": "GitHub", "url": "https://github.com" } } ]