… roles

…out kwargs

Update all run_llm_benchmark_cli() calls in the integration test file from the
retired output_path/checkpoint_path/artifacts_dir kwargs to output_dir/run_id,
and read persisted artifacts back from the RunLayout-derived paths
(run_dir/summary.json, predictions_dir, cases.jsonl) instead of the old
single output_path file. Covers 5 tests flagged by review plus 4 more in the
same file with the same root cause that were missed because they don't match
"llm" in their test names.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…tifacts-dir with --output-dir/--run-id/--overwrite

…/llm runs

…ifests

terms.jsonl/cases.jsonl are only computed once at completion, so a
partial-run manifest's counts.terms/counts.cases previously always
read 0 while a checkpoint was still in progress.

Run Artifacts and LLM Full-Text Benchmark sections still described the
pre-unification results/llm/<timestamp>.json layout and a stale output
JSON key list.

… matches

_build_benchmark_payload never wrote capture_phase1_debug into the
checkpoint, so _checkpoint_matches_run compared the stored default
(None) against the CLI's real default (False) and always rejected the
checkpoint on any real --run-id/--overwrite resume. Discovered live
while resuming an interrupted OpenRouter run.

…/base_url

current_run compared the raw --llm-provider/--llm-base-url CLI inputs,
but run_llm_benchmark() persists the provider resolver's resolved
values (e.g. the default provider when --llm-provider is omitted, or
the model-prefix-inferred provider). Resolve once in the CLI so a
resume with the same omitted flags matches instead of always rejecting
the checkpoint.

Checkpoints written before capture_phase1_debug started being
persisted still lack the key entirely, so a resume against one of
those legacy files failed the same match check the previous fix
addressed for newly-written checkpoints. Discovered resuming a real
interrupted run whose checkpoint predated that fix.

…tory with 8 updates

Bumps the frontend-minor-patch group with 6 updates in the /frontend directory:

| Package | From | To |
| --- | --- | --- |
| [libphonenumber-js](https://gitlab.com/catamphetamine/libphonenumber-js) | `1.13.7` | `1.13.8` |
| [@vitest/coverage-v8](https://github.com/vitest-dev/vitest/tree/HEAD/packages/coverage-v8) | `4.1.9` | `4.1.10` |
| [eslint](https://github.com/eslint/eslint) | `10.6.0` | `10.7.0` |
| [postcss](https://github.com/postcss/postcss) | `8.5.16` | `8.5.18` |
| [prettier](https://github.com/prettier/prettier) | `3.9.4` | `3.9.5` |
| [terser](https://github.com/terser/terser) | `5.48.0` | `5.49.0` |



Updates `libphonenumber-js` from 1.13.7 to 1.13.8
- [Changelog](https://gitlab.com/catamphetamine/libphonenumber-js/blob/master/CHANGELOG.md)
- [Commits](https://gitlab.com/catamphetamine/libphonenumber-js/commits/master)

Updates `@vitest/coverage-v8` from 4.1.9 to 4.1.10
- [Release notes](https://github.com/vitest-dev/vitest/releases)
- [Changelog](https://github.com/vitest-dev/vitest/blob/main/docs/releases.md)
- [Commits](https://github.com/vitest-dev/vitest/commits/v4.1.10/packages/coverage-v8)

Updates `@vitest/ui` from 4.1.9 to 4.1.10
- [Release notes](https://github.com/vitest-dev/vitest/releases)
- [Changelog](https://github.com/vitest-dev/vitest/blob/main/docs/releases.md)
- [Commits](https://github.com/vitest-dev/vitest/commits/v4.1.10/packages/ui)

Updates `eslint` from 10.6.0 to 10.7.0
- [Release notes](https://github.com/eslint/eslint/releases)
- [Commits](eslint/eslint@v10.6.0...v10.7.0)

Updates `postcss` from 8.5.16 to 8.5.18
- [Release notes](https://github.com/postcss/postcss/releases)
- [Changelog](https://github.com/postcss/postcss/blob/main/CHANGELOG.md)
- [Commits](postcss/postcss@8.5.16...8.5.18)

Updates `prettier` from 3.9.4 to 3.9.5
- [Release notes](https://github.com/prettier/prettier/releases)
- [Changelog](https://github.com/prettier/prettier/blob/main/CHANGELOG.md)
- [Commits](prettier/prettier@3.9.4...3.9.5)

Updates `terser` from 5.48.0 to 5.49.0
- [Changelog](https://github.com/terser/terser/blob/master/CHANGELOG.md)
- [Commits](terser/terser@v5.48.0...v5.49.0)

Updates `vitest` from 4.1.9 to 4.1.10
- [Release notes](https://github.com/vitest-dev/vitest/releases)
- [Changelog](https://github.com/vitest-dev/vitest/blob/main/docs/releases.md)
- [Commits](https://github.com/vitest-dev/vitest/commits/v4.1.10/packages/vitest)

---
updated-dependencies:
- dependency-name: libphonenumber-js
  dependency-version: 1.13.8
  dependency-type: direct:production
  update-type: version-update:semver-patch
  dependency-group: frontend-minor-patch
- dependency-name: "@vitest/coverage-v8"
  dependency-version: 4.1.10
  dependency-type: direct:development
  update-type: version-update:semver-patch
  dependency-group: frontend-minor-patch
- dependency-name: "@vitest/ui"
  dependency-version: 4.1.10
  dependency-type: direct:development
  update-type: version-update:semver-patch
  dependency-group: frontend-minor-patch
- dependency-name: eslint
  dependency-version: 10.7.0
  dependency-type: direct:development
  update-type: version-update:semver-minor
  dependency-group: frontend-minor-patch
- dependency-name: postcss
  dependency-version: 8.5.18
  dependency-type: direct:development
  update-type: version-update:semver-patch
  dependency-group: frontend-minor-patch
- dependency-name: prettier
  dependency-version: 3.9.5
  dependency-type: direct:development
  update-type: version-update:semver-patch
  dependency-group: frontend-minor-patch
- dependency-name: terser
  dependency-version: 5.49.0
  dependency-type: direct:development
  update-type: version-update:semver-minor
  dependency-group: frontend-minor-patch
- dependency-name: vitest
  dependency-version: 4.1.10
  dependency-type: direct:development
  update-type: version-update:semver-patch
  dependency-group: frontend-minor-patch
...

Signed-off-by: dependabot[bot] <support@github.com>

…produce

Two defects in the shared result store, both of which undermine the
reproducibility guarantee the store exists to provide.

`create_run_layout(overwrite=True)` reused a run directory without clearing it.
`write_manifest` decides what to advertise purely from `Path.exists()`, so a
rerun under the same `--run-id` inherited the previous run's files and published
them as its own: re-running an extraction without diagnostics still advertised
the earlier `diagnostics/chunks.jsonl`, and a narrower `--doc-ids` LLM rerun left
the earlier documents inside the advertised `predictions/` directory. Clear the
run directory on overwrite, preserving only `checkpoint.json` — the LLM benchmark
reuses a run directory precisely to resume from it, and that checkpoint's
identity is validated separately before any of it is trusted. The checkpoint
filename is now owned by `RunLayout` instead of being duplicated as a literal.

`discover_artifacts` was also type-blind. All three benchmark types write a
`summary` role into the shared results root, so the recursive discovery handed
extraction and LLM summaries to the two loaders that render dense-retrieval
tables. Those summaries carry no `mrr_dense`, so `phentrieve benchmark compare`
listed them as models scoring `MRR (Dense) = 0`, and every new extraction or LLM
run added another junk row. Add an optional `benchmark_type` filter and set it in
both retrieval loaders; the legacy fallback stays unfiltered because the old flat
layout only ever held retrieval summaries.

`load_summary_files` also swapped `glob.glob` (str) for `discover_artifacts`
(Path) without updating the `file_path` it injects into each summary, which broke
JSON serialization of loaded summaries.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GfjxrHsHYH2TqjwgEcq6oS

`cases.jsonl` mixed two identifier schemes. Evaluable and failed cases were
keyed on a 0-based `case_id` — the same key `terms.jsonl` joins on — while cases
skipped for having no gold terms were keyed on a 1-based `case_index` and carried
no `case_id` at all. A downstream join over the artifact therefore silently
dropped exactly the rows the manifest counts under `not_evaluable`, and the two
index bases sat in one file.

Emit `case_id` (and the case text) for non-evaluable rows too, so every row in
`cases.jsonl` is joinable on one stable key.

Also return `run_dir` as a str: it is part of a public result dict, and the LLM
payload already returns a str, so a caller serializing the retrieval result hit
`TypeError: Object of type PosixPath is not JSON serializable`.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GfjxrHsHYH2TqjwgEcq6oS

…tatus

The checkpoint identity recorded `prompt_templates_dir` but not the templates
themselves. The prompt is an experimental variable, so editing a template in
place between two runs of the same `--run-id` left the identity unchanged and let
the resume merge old-prompt and new-prompt document outputs into one set of
metrics. Hash the templates a run will actually load — both the packaged set and
the effective user set, which is `~/.phentrieve/prompts` even when the flag is
omitted — and make that hash part of the identity.

Checkpoints written before the key existed carry no evidence about the templates
they ran under, so they stay resumable under the behaviour they were written with
rather than being invalidated on upgrade; checkpoints written from now on always
carry the hash and are compared strictly.

The manifest also always claimed `status: complete`. `run_llm_benchmark` reports
`completed` for any run that finishes its document loop, including one in which
every document failed, and the CLI mapped anything not equal to `failed` onto
`complete`. Derive the run status from the per-case outcomes instead — as the
retrieval runner already does — and record failed/complete counts so the status
is auditable.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GfjxrHsHYH2TqjwgEcq6oS

Both new artifact writers open-coded `1.0 - distance` when the retriever returned
distances without similarities, while the retriever itself converts through
`calculate_similarity`, which clamps to [0, 1]. An unclamped score also feeds
`passes_threshold` in the extraction chunk diagnostics. Use the shared helper so
the artifacts agree with the scores the pipeline actually acted on.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GfjxrHsHYH2TqjwgEcq6oS

The guide told users to pass `--summaries-dir results/retrieval` to work around
comparisons picking up non-retrieval runs; the loaders now filter by benchmark
type, so the plain results root is safe.

It also described `traces/` as written only with `--capture-phase1-debug`. The
prediction record always carries a trace, so traces are always written; the flag
adds phase-1 retrieval detail to them rather than enabling them.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GfjxrHsHYH2TqjwgEcq6oS

…/vite-8.1.4' into chore/benchmark-result-store

Merges the two open Dependabot frontend PRs into this branch. Both edit the same
dependency block, so package.json is resolved as the union of the two — vite
^8.1.4 from #307 alongside the six bumps from #308 (libphonenumber-js, eslint,
postcss, prettier, terser, @vitest/coverage-v8) — and package-lock.json is
regenerated from the resolved manifest rather than hand-merged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GfjxrHsHYH2TqjwgEcq6oS

CLI takes a minor bump for the benchmark result store and its new flags; the
frontend takes a patch bump for the folded-in dependency updates. The API is
untouched by this change and keeps 0.17.0.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GfjxrHsHYH2TqjwgEcq6oS