feat: v0.4.0 — rich content support with typed blocks and loss visibility

Extracts per-message content into a typed `blocks` list (text, code,
thinking, tool_use, tool_result, image_placeholder, file_placeholder,
unknown) and renders them at exporter write time. Voice transcripts,
Custom Instructions, and image references now appear in exports
instead of being silently dropped.

Foundation:
- src/blocks.py: pure block constructors, _safe_fence (fence-corruption
  defense, verified live in Joplin), _blockquote_prefix, render
- src/loss_report.py: per-run tally surfaced as INFO summary at end of
  export so silently-dropped data becomes visible

Providers:
- ChatGPT: dispatch on content_type produces typed blocks; voice shapes
  (audio_transcription, audio_asset_pointer, real_time_user_audio_video_
  asset_pointer) locked from live DevTools capture; Custom Instructions
  bug fix (parts-vs-direct-fields); role filter lifted; hidden-context
  marker driven by is_visually_hidden_from_conversation flag
- Claude: defensive dispatch for text/thinking/tool_use/tool_result/image
  with recursive nested-block flattening; untested against real rich-
  content data — fix-forward in v0.4.1

Exporter:
- Markdown renders from blocks at write time via render_blocks_to_markdown;
  backward-compat fallback to content for any pre-v0.4.0 cached data

Tests:
- 27 new tests across providers, exporters, CLI; fixtures rebuilt with
  real-shape ChatGPT voice + Custom Instructions cases
- 181/181 pass

Behavior changes (intentional):
- JSON output omits content; consumers should read blocks
- Per-conversation message counts increase (Custom Instructions, image-
  only, tool-only messages now appear)
- Existing exports not auto-re-rendered; users wanting fresh output run
  cache --clear then export

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

CHANGELOG.md

@@ -3,6 +3,41 @@
 All notable changes to this project will be documented here.
 Format follows [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
 
+## [0.4.0] - Unreleased
+### Added
+- Rich content support: messages now carry an ordered `blocks` list (text, code, thinking, tool_use, tool_result, citation, image_placeholder, file_placeholder, unknown)
+- ChatGPT voice mode: `audio_transcription` parts render as text blocks; `audio_asset_pointer` and `real_time_user_audio_video_asset_pointer` render as `📎 File attached` placeholders with size and duration metadata
+- ChatGPT Custom Instructions: `user_editable_context` and `model_editable_context` messages now appear in exports (were silently dropped — pre-existing bug fixed); rendered with a `> ℹ️ Hidden context` marker driven by the `is_visually_hidden_from_conversation` flag
+- Image placeholders for `image_asset_pointer` parts (uploads + DALL-E) inside `multimodal_text` and at message level
+- Defensive Claude block extraction: `text`, `thinking`, `tool_use`, `tool_result` (including nested-block flattening), `image` blocks (untested against real data; will fix-forward in v0.4.1 if real shapes diverge)
+- `LossReport` summary table emitted at end of every `export` run, breaking down `unknown blocks` and `extraction failures` by raw type so silently-dropped data becomes visible
+- `_safe_fence` helper picks a fence longer than any backtick run in extracted content, preventing embedded triple-backticks from corrupting downstream rendering (verified live in Joplin during planning)
+- `unknown` blocks render as `> ⚠️ Unsupported content` with the raw type, observed top-level keys, and reason — so future API additions are visible rather than silent
+
+### Changed
+- ChatGPT role filter (previously dropped `tool` and `system` messages) is **lifted**: all roles now route through normal extraction; truly empty messages skip via the existing empty-content guard
+- Markdown rendering moves from provider-time to exporter-write-time. Providers produce blocks; exporters call `render_blocks_to_markdown` at write time. This unblocks future Obsidian/HTML exporters
+- `BaseProvider.normalize_conversation` signature now accepts an optional `LossReport` parameter (breaking change for any future custom subclass; FileProvider hasn't shipped yet)
+- `o1`/`o3` reasoning subparts inside `text` content_type messages remain rendered as plain text (defensive; reclassification to `thinking` block deferred until live shape is captured)
+
+### Fixed
+- `user_editable_context` / `model_editable_context` extraction (parts-vs-direct-fields mismatch) — Custom Instructions are no longer silently dropped from every conversation
+
+### Migration
+- Existing exports are not re-rendered automatically. To pick up v0.4.0 rendering for previously exported conversations:
+  ```
+  python -m src.main cache --clear
+  python -m src.main export --provider all
+  ```
+- JSON exports: messages now contain `blocks` (typed structured content) and may omit the legacy `content` field. External consumers reading JSON should prefer `blocks`.
+- Per-conversation message counts may increase: previously-dropped Custom Instructions, image-only user turns, and tool-only assistant turns now appear.
+
+### Out of scope (deferred to v0.5.0+)
+- Binary downloads of images and audio assets (placeholders show metadata only; `content not preserved in this export`)
+- Joplin resource upload for embedded media
+- Filename resolution for `file-XYZ` / `sediment://` references
+- Speculative ChatGPT types (`tether_browsing_display`, `tether_quote`) and DALL-E assistant images — fall through to `unknown` blocks if encountered
+
 ## [0.2.0] - Unreleased
 ### Added
 - Joplin import automation: `joplin` command syncs exported Markdown files to Joplin as notes