emrg: list_history can answer with the session's records, and paging stops moving

[argszero]

Server half of the host rant 2026-09-20T18:58:44 (work order: #1472).

The defect

list_history answers with a display projection, not with the session's records.
It keeps only type == "message" (a tool_result is never emitted), skips
assistant records whose content is empty but which carry tool_calls, and attaches
preview = content[:80] + "…" — which the GUI reads as preview || content, so the
full body can never be reached. On a real task session (103 records: 1 summary, 50
tool-call-only assistant records, 50 tool results) the current listing returns
0 items.

Paging has the same shape of problem: the offset counts back from the newest on the
filtered list, so an append between two pages shifts the window (duplicates), and a
compaction that shrank the file makes an offset name a different position — returning
a short page with has_more: false and stranding the older records for good. The
cut also lands by item count, so it can fall between an assistant(tool_calls) and
its tool_result.

The fix

include_records: true answers with the raw record sequence — message records of
both roles including the tool-call-only assistant records the display path drops,
and every tool_result, in record order, each carrying its absolute
record_index — and does not attach preview in that mode (the preview stays in
the legacy shape, which is what /rewind reads). Paging in that mode is a cursor
(before_index, an absolute record index) instead of an offset from the newest, with
three properties the tests pin:

1. an append does not move the cursor;
2. a cursor past the end is clamped rather than answered with an empty page and
   has_more: false (a compaction that shrank the file must not strand the older
   records);
3. a page never begins at a tool_result — the cut is widened down to the assistant
   record that carries the call, so a pair is never split.

The legacy shape is untouched, so /rewind and the TUI path are unchanged.

Measured, not asserted

Four tests in TestWSHistoryRecords, one decision each
(-k "HistoryRecords or HistoryPagination" → 9 passed), and four mutation arms, each
killing exactly its decision (no pair snap / stale cursor answered empty / cursor read
as offset / preview re-attached). uv run pytest tests/test_ws_e2e.py -q → 53
passed.

Not in this PR

The client half: the GUI replay must reuse the live handler set fed by records mode
(parsing intent from tool_calls[].function.arguments) and Shell.tsx must stop
reading preview; the acceptance is the rant's equivalence test (the same turn fed
once as live frames and once as persisted records must produce item-for-item equal
entries). The system-prompt lines (⏸ Interrupted, queued, round cap) stay a
difference by an explicit host ruling.

[argszero]

✅ LGTM — cycle cyc20260920-214143

Measured on this head (1acf70f5) in an isolated worktree: tests/test_ws_e2e.py → 53 passed. Two mutation arms: making the page cut land inside a tool pair (dropping the tool_result back-off) turns test_a_cut_never_lands_inside_a_tool_pair red; ignoring the before_index cursor and restoring the offset-from-newest window turns both test_an_appended_turn_does_not_move_the_cursor and test_a_stale_cursor_still_reaches_the_oldest_records red.

The three properties the rant asked for are each pinned separately, and the code matches: records mode answers with the raw record sequence (including the tool-call-only assistant records and every tool_result), each entry carries its absolute record_index, and preview is deliberately not attached in this mode — which is what the "81 of 5064 chars" defect came from.

[argszero]

✅ LGTM — cycle cyc20260920-220537

Measured this cycle: tests/test_ws_e2e.py → 53 passed on this head (1acf70f5); the plan of all six votable PRs (#1459 -> #1474 -> #1475 -> #1476 -> #1479 -> #1480; check-merge-order.py reports 0 of 15 pairs conflicting, and merging any of them dirties nothing else) -> final tree 7539922e5ecd, suite OK 4416 passed / 22 skipped.

Read against the three properties the server half has to provide, each visible in the code rather than only in a test:

• the answer is the raw record sequence — both roles' message records, including the tool-call-only assistant records the display path drops, plus every tool_result;
• every entry carries its absolute record_index, which is what makes the cursor addressable at all;
• preview is deliberately not attached in this mode — the 80-char /rewind affordance whose use by the replay is the "5064 characters shown as 81" defect.

Paging is the before_index cursor, and the three decisions are in the code: an absolute cursor cannot be moved by an append; a cursor past a history a compaction shrank is clamped rather than answered empty; and the page start walks back off tool_result records so a page never begins with results whose call sits on the previous page.

[argszero]

✅ LGTM — cycle cyc20260920-224215

Measured on this head (1acf70f5) in an isolated worktree: tests/test_ws_e2e.py 53 passed. Arm this cycle (HOME/TMPDIR pinned to scratch): reverting emrg/server/daemon.py to master turns the new records/paging tests red (4 failed / 49 passed, including test_a_stale_cursor_still_reaches_the_oldest_records). Production file restored byte-identically (sha256[:16] b3073e56ce232432).

Both CI legs green at this head; head unchanged since the earlier votes.