ci: index the docs on Context7 and refresh on push

[cardmagic]

Why

Context7 has no classifier library yet, so coding agents that use it cannot retrieve the current API.

What changes

• context7.json limits indexing to the README and docs/, leaves out AGENTS.md, CLAUDE.md, and CONTRIBUTING.md, and gives agents five rules taken from the docs. It also carries the url and public_key that Context7 reads to verify the maintainer's claim. The public_key is the maintainer's public account key, which Context7 designs to be committed.
• .github/workflows/context7-refresh.yml asks Context7 to refresh the index after a push to master that changes the README, context7.json, or docs/. Context7 allows one refresh every 10 days for a library and answers too-early before then. That answer ends the job with a notice. Any other non-200 answer fails the job. It reads the CONTEXT7_API_KEY repository secret, which is set.

No CHANGELOG entry: the change does not reach gem users.

Validation

• actionlint passes, and context7.json parses.
• Each rule was checked against the docs and code: the class names exist, docs/tfidf.md documents Classifier::NotFittedError, docs/persistence.md says to prefer JSON over Marshal, and train_from_stream and train_batch exist.
• The same workflow script runs in cardmagic/solid-objects-ruby and cardmagic/solid-objects-js, where it passed against the live API.

[greptile-apps]

Confidence Score: 5/5

[High risk] Adds CI workflow that calls external API with a secret.

The PR appears safe to merge, with one non-blocking correction to the TF-IDF guidance.

Findings

1.  TF-IDF guidance omits memory limits ▶

Fix with agent prompt

### Issue 1
context7.json:20
The rule now presents `fit_from_stream` as the `Classifier::TFIDF` option for data larger than memory. However, it collects every document in an array before calling `fit`. Agents following this advice can generate code that runs out of memory on large corpora. State that `fit_from_stream` still requires the full corpus to fit in memory.

```suggestion
    "For data larger than memory, use train_from_stream, which reads one document per line from an IO stream. Classifier::TFIDF uses fit_from_stream, but it collects all documents in memory before fitting, so the corpus must fit in memory. Use train_batch only for documents already in memory; Classifier::LSI and Classifier::KNN name it add_batch."
```

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Adds Context7 indexing settings and a workflow that requests a refresh after documentation changes on master.

• The latest commit fixes the previous advice about train_batch.
• The new TF-IDF advice still needs a note that fitting retains the whole corpus in memory.

_{Reviews (2) · Last reviewed commit: "ci: describe batch training as in-memory" · Reviewed by Greptile}

[cardmagic]

@greptileai Please review the latest commit 4f8e225. The streaming rule now matches docs/streaming.md, including fit_from_stream for TF-IDF and add_batch for LSI and KNN.

[greptile-apps]

TF-IDF guidance omits memory limits

The rule now presents fit_from_stream as the Classifier::TFIDF option for data larger than memory. However, it collects every document in an array before calling fit. Agents following this advice can generate code that runs out of memory on large corpora. State that fit_from_stream still requires the full corpus to fit in memory.

Suggested change

	"For data larger than memory, use train_from_stream, which reads one document per line from an IO stream. Classifier::TFIDF uses fit_from_stream. Use train_batch only for documents already in memory; Classifier::LSI and Classifier::KNN name it add_batch."
	"For data larger than memory, use train_from_stream, which reads one document per line from an IO stream. Classifier::TFIDF uses fit_from_stream, but it collects all documents in memory before fitting, so the corpus must fit in memory. Use train_batch only for documents already in memory; Classifier::LSI and Classifier::KNN name it add_batch."

Knowledge Base Used: Streaming processing

Prompt To Fix With AI

This is a comment left during a code review.
Path: context7.json
Line: 20

Comment:
**TF-IDF guidance omits memory limits**

The rule now presents `fit_from_stream` as the `Classifier::TFIDF` option for data larger than memory. However, it collects every document in an array before calling `fit`. Agents following this advice can generate code that runs out of memory on large corpora. State that `fit_from_stream` still requires the full corpus to fit in memory.

```suggestion
    "For data larger than memory, use train_from_stream, which reads one document per line from an IO stream. Classifier::TFIDF uses fit_from_stream, but it collects all documents in memory before fitting, so the corpus must fit in memory. Use train_batch only for documents already in memory; Classifier::LSI and Classifier::KNN name it add_batch."
```

**Knowledge Base Used:** [Streaming processing](https://app.greptile.com/craftsmanfounder/-/custom-context/knowledge-base/cardmagic/classifier/-/docs/streaming-processing.md)

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.