Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@src/metrics.rs`:
- Around line 13-31: The DuckDB refresh is too weak because deletion/refresh
logic only keys off app_id and period bounds, allowing mixed rollups to
conflict; add persistent fetch-shape discriminator columns to the metrics table
(e.g., interval, group_by, metric_set or similar column names used in your
ingest path) and change the refresh/delete logic in the refresh routine to
include those discriminator columns when deleting/upserting rows for a staged
load (replace the current app_id + MIN/MAX(period_*) logic). Also ensure the
refresh path (the code that computes staged MIN/MAX(period_start/period_end))
rejects empty staged fetches or falls back to a safe sentinel so NULL MIN/MAX
cannot cause no-op deletions; if you prefer, alternatively make db writes reject
mixed-shape writes by validating the staged fetch-shape against existing stored
discriminator values before deleting/upserting.

---

Nitpick comments:
In `@skills/apitally-cli/references/commands.md`:
- Around line 84-144: The markdown code fences for the `metrics` CLI docs are
missing language identifiers; update the three new fences so the first command
block (the npx `@apitally/cli` metrics example) is tagged as bash, and the two
JSON example blocks (the Filters examples block and the Example NDJSON output
block) are tagged as json—look for the triple-backtick blocks surrounding the
npx command, the filter examples JSON array, and the NDJSON output and add the
appropriate language identifiers.

In `@skills/apitally-cli/SKILL.md`:
- Around line 59-144: The code fences in SKILL.md for CLI commands (e.g., the
npx `@apitally/cli` examples such as the endpoints, consumers, metrics,
request-logs, request-details commands) and the SQL examples (the DuckDB queries
under "Query DuckDB" and the investigation SQL blocks) need explicit languages
to satisfy MD040; update each backtick-fenced block so command examples use
```bash and SQL examples use ```sql (search for snippets containing npx
`@apitally/cli` and the SELECT ... FROM request_logs / metrics queries) to remove
the lint warnings.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@skills/apitally-cli/references/commands.md`:
- Around line 86-90: Add a language tag (e.g., bash) to the fenced code block
that contains the npx `@apitally/cli` metrics usage example so the markdown fence
becomes ```bash ... ```, updating the block that starts with "npx `@apitally/cli`
metrics <app-id> --since <datetime> --metrics <json> \" to include the language
identifier for proper linting and rendering.
- Around line 103-104: Update the "Deduplication in DuckDB" description to
precisely reflect the implemented bounded delete semantics: state that for each
app_id the process deletes existing rows where period_start >= min(period_start)
and period_end <= max(period_end) of the staged (fetched) rows, rather than
saying "within the fetched time range", so readers understand the delete uses
those exact bounds and not arbitrary overlap matching.