[FEATURE] Granular control (turn limit / redact) over conversation context attached to feedback

### Preflight Checklist

- [x] I have searched [existing requests](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20label%3Aenhancement) and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)

### Problem Statement

The new structured `/feedback` / feedback-submission format (labeled fields: observed behavior, the user's own words, repro steps, named evidence) is a substantive improvement over a scalar rating, and the reason is worth stating precisely, since it bears directly on the gap below.

A scalar rating (1-3 stars, thumbs up/down) is unfalsifiable: it asserts no checkable claim about the interaction — nothing about what happened, what was expected, or what would make it wrong. This has three compounding consequences:

1. **No inter-rater reliability.** Two raters observing the same output are not disagreeing about a shared fact, because the scale contains no fact to agree or disagree about; scores do not converge toward anything measurable.
2. **Selection bias.** Users experiencing friction rate substantially more often than satisfied users, so the aggregate tracks the distribution of frustration, not the distribution of quality.
3. **Unaccountable attribution.** Most raters cannot distinguish a genuine model error from a tool failure, an underspecified request, or a defensible judgment call they simply disagreed with. Because the score is unfalsifiable, that attribution can never be checked or contested after the fact.

The combined effect is an information-laundering mechanism: a signal with no verifiable epistemic content is nonetheless aggregated, reported, and acted on downstream as though it were evidence of model quality. The structured feedback format avoids this by construction — every field is a claim that could in principle be shown false, which is what makes it usable as evidence rather than noise.

What hasn't caught up is the context-attachment control. Today it is last-session / 24h / 7d / none (per #89874) — all-or-nothing. For a user working across confidential codebases at high volume, "none" discards exactly the context that makes the structured format actionable, while "last session" can mean hours of unrelated material archived to substantiate one narrow report.

### Proposed Solution

Let the user scope the attached context precisely rather than choosing all-or-nothing:

- A "last N turns" option (a number, not a time window), so exactly the turns relevant to the report are attached.
- A preview-and-redact step before submission, so specific lines/turns can be stripped without discarding the whole transcript.

This is not purely a privacy request: the current binary choice is itself suppressing feedback volume from the users best positioned to supply well-evidenced reports.

### Alternative Solutions

The existing "none" option (#89874) is a valid but blunt fallback — it solves the privacy problem by discarding exactly the evidence that makes a structured report actionable. It's a workaround, not a fix for the underlying scoping gap.

### Priority

High - Significant impact on productivity

### Feature Category

CLI commands and flags

### Use Case Example

1. I'm 90 minutes into a long working session across several unrelated tasks in a confidential codebase.
2. Near the end, I hit a specific, well-defined bug in a CLI permission classifier.
3. I want to file a precise, evidence-backed report on that one issue — but the only context-attachment options are "last session" (which includes 90 minutes of unrelated confidential work) or "none" (which discards the very evidence the structured format is designed to carry).
4. With "last N turns," I could attach just the handful of turns that actually demonstrate the bug.

### Additional Context

Related: #89874 (session-history-optional request), #89130 (unnecessary git repo data attached to feedback). This issue asks for precision *within* the "with context" branch, rather than only an on/off toggle.

Companion issue, filed alongside this one: a request to let users view/track feedback they've submitted (see linked issue below once filed).

**If this is addressed, I will make it worth your while: once my privacy needs are met, I commit to filing feedback daily, at the highest volume I can sustain.**

Start with the existing /feedback flow and its current last-session/24h/7d/none context choices described here and in #89874. Completion means supporting a last-N-turns scope and preview/redaction before submission while preserving the existing no-context option; the payload names no implementation files or tests.

Scritto dal modello di indicizzazione a partire dal testo della issue.