Generate input echoes from sideband events

[t-kalinowski]

Summary

This PR removes output-inspection echo handling and makes consumed-input sideband events the only source of generated input echo text. The point is to stop treating prompt-shaped stdout/stderr as something to infer or suppress, while still preserving transcript context in overflow surfaces that need it.

Normal MCP replies now hide generated input echoes. Files-mode bundle transcripts and pager overflow pages keep generated echoes from the resolved timeline, so transcript-style surfaces still show consumed input without inspecting captured output.

User-facing changes

• Normal repl replies no longer include generated input echoes; prompt-shaped stdout/stderr is preserved only when it was real output.
• Files-mode output bundles include full generated input echoes in transcript.txt, even when the inline MCP preview omits them.
• Pager overflow pages include full generated input echoes as separate text blocks before raw output.
• Timeout follow-up prefixes and other non-overflow pager drains do not expose generated input echoes.

Internal changes

• Added src/resolved_output.rs as the shared renderer for raw bytes, sideband events, text spans, images, and generated input echoes.
• Deleted the pager-specific merge path and removed output-based echo trimming/suppression helpers.
• Added ContentVisibility so generated echoes can be transcript-only without being serialized into normal MCP replies.
• Kept files and pager backing stores separate: files mode still owns long-lived transcripts, request sealing, and timeout staging; pager mode still owns bounded ring, page cursors, and search state.
• Updated docs, integration expectations, unit tests, and snapshots for the generated-echo contract.

Notes

• R/Python input() and readline() normal-reply echo presentation remains follow-up scope; overflow transcript surfaces preserve generated input echo.
• This remains a draft PR.

Diff composition

Measured against main, this PR is 1505 insertions and 4114 deletions across 66 files.

• runtime src/: +901/-2774 (65.4% of churn)
• inline tests inside src/: +191/-1121 (23.3% of churn)
• tests in tests/: +252/-144 (7.0% of churn)
• docs: +49/-53 (1.8% of churn)
• snapshots: +112/-22 (2.4% of churn)

Largest files:

• src/output_timeline.rs: +0/-1262
• src/reply_presentation.rs: +26/-637
• src/pending_output_tape.rs: +131/-509
• src/resolved_output.rs: +455/-0
• src/server/tests.rs: +0/-343