refactor(web,server): match a turn to its reply by execution id, and clamp text through one primitive

[gbasin]

The two follow-ups worth doing from the #538/#539 thread, built in parallel lanes and reviewed here. The third (tokenizing WorkFold's ml-[43px]) I dropped after investigating — see the bottom.

1. Fold→reply matching by identity, not position

session.replied carried {session_id, text, broadcast} and no turn identity, so buildSpineRows matched by ordinal: the Nth reply-producing fold in the SSE stream was assumed to be the Nth session.replied in the thread. True only while they stay 1:1 — and a failed execution posts no session.replied, so the two clocks drift and a turn's work can nest under someone else's answer. (Prod thread 575 has exactly this shape: 5 executions, 4 replies. It doesn't currently mis-attribute only because the failed turn produced no fold.)

The terminal execution_state frame that drives completeSession already knows which execution answered, so it now says so:

• server — session.replied payload carries execution_id
• reducer — SessionState tracks the current execution; every item is stamped with the execution that produced it (following the existing Artifact.executionId precedent)
• folds — FoldedTurnRow.executionId
• wire — ChatMessage.sessionExecutionId, decoded-with-default so legacy events yield null rather than throwing
• placement — matches on identity; the ordinal machinery is deleted

No fallback, deliberately (product call): replies written before this change have no execution id and simply don't nest — their work keeps a standalone row. One code path, paid for with nesting on existing threads. Where a steer splits a running execution into several folds, the execution's answer is the last one — that nests, earlier ones stay standalone. No fold is ever dropped.

2. One clamp primitive

#538's bug was two clamps nesting: CompactReply's line-clamp-3 wrapped MessageText's own max-h-80, so "Show more" shrank the message (360px → 87px in a real browser). #538 fixed it with a collapsible opt-out — a convention, and conventions get forgotten.

ClampedBlock puts the clamp in React context, so an inner block declines to clamp: the nesting is now unrepresentable, not merely discouraged. It also measures overflow rather than trusting a length heuristic, so a message that trips the threshold but fits on screen no longer offers a toggle that does nothing (the one visible behaviour change).

Adopted at 2 of 4 sites, on purpose. EntryQuoteCard collapses a changes list behind a domain-specific action row; SessionCapabilitiesPopover slice()s an item array and has no overflowing box to measure. Forcing either through the primitive would bend it out of shape for no benefit — the duplication there is honest.

Verification

• 729 web · 171 centaur-client · 428 shared · 893 server ✓ · biome ✓ · typecheck ✓
• Real browser: cluster-steer-interleave (incl. "Show more grows a long agent reply" — jsdom has no layout and cannot judge it) and session-pane-ux work-fold specs all pass through the new primitive.
• Local full-suite e2e is flaky on my machine right now and CI is the arbiter. I baselined it: stock master with zero changes fails 1–2 tests per full run (addressability-shell:81, nav-discoverability:23), a rotating cast that all pass in isolation. This branch shows the same rate and character. I nearly mis-attributed that to these changes.

Dropped: the WorkFold gutter token

ml-[43px] / calc(100%-59px) appear exactly once in the codebase, already behind a nested prop with an explanatory comment. There's no duplication to remove, and the numbers don't align to anything I could name truthfully (MessageRow's content column computes to 60px, not 43). Naming a single-use constant would be indirection for its own sake.