W5b-14: timing.Ledger wired into the engine stages — one measurement per interval

[bnsd55]

Per scratchpad/w5b-timing-ledger.md and the required-parameter design: ONE jevmlx.timing.Ledger per request, passed DOWN as a required parameter — no optional/dual path.

• Spans: plan, prefill, cache_merge/ transformer/ gather (inside _score_rows, per chunk), rescore (field finalization + batch=1 rescores), reconciliation (constraint MAP), dependency (the selective second pass), assembly (batched per-context), group_wall (batched per group); the neutral prior pass is a prior-phase span.
• ScoreRowsResult.gather_ms/broadcast_ms deleted; the internal t_gather_ms/t_broadcast_ms accumulators gone. A failed forward/gather records no interval (the ledger drops spans an exception unwinds through).
• finalize_public_result derives EVERY *_ms key from Ledger.derived_flat() — same contract keys (suffix_eval_ms stays the documented composite), no accumulator fallback.
• Batched (GAP A fix): ONE Ledger PER CONTEXT — each result's flat *_ms keys derive from its OWN ledger (own prefill span + own assembly spans + amortized share of the group's merged scoring pass); group_wall_ms/per_item_amortized_ms come from the per-group ledger; per_item_end_to_end_ms = the context's own prefill-span start → its assembly-span end (the note's definition; the shared prior_ms is reported separately, not added).
• check_results/ARCHITECTURE/CHANGELOG updated; keys unchanged.

Tests: 7 new (ledger-derivation consistency, prior phase, dependency span, group-vs-amortized, per-item e2e, no dual fields). 697 fast passed; slow suite green except the pre-existing api.py '#count' KeyError — routed to PR #51 by the coordinator. Rebased over 4a656e6 (includes #54's prior-cache-key fix).

Review round 1 (77af115)

• GAP B: Ledger.__exit__ with an exception unwinds ONLY to the failing span — ancestors stay open (a Metal retry under group_wall no longer drops the parents); nested spans are dropped.
• F1 conflict marker resolved; F2 dual path + t0 deleted; F3 second_pass_ms = dependency span only; F6 _eval_cache_state unconditional; F7 R==0 branch spans fixed; F8 transformer span covers mx.eval(out); F9 leftovers deleted (PrefillResult.t_prefill_ms, stale docstrings, dead timers); F10 one top-level request span (elapsed_ms = true wall); F11 with everywhere except _score_rows retry paths; F12 module-top import, prior_pass name, cache-hit note; F13 body/code agree.
• New tests: per-context prefill_ms differ; second_pass_ms == dependency span; forced Metal retry under group_wall; unpadded cache eval unconditional. 721 passed (slow suite green except the pre-existing api.py KeyError routed to PR W5-C: exact setcons solver, one count+set optimization, typed CalibrationBundle, internal telemetry, abstention contract #51).

Review round 3 (d1e9dd4, rebased on 5449b64)

• N2 crash fixed: the R==0 branch shares the main loop's store path (no nested group_wall, no double pop); repro is a test.
• N1 fixed: every batched result reports the shared prior_ms (derived_flat(prior_ms=...) override for per-context ledgers); total_ms includes it; test added.
• REAL BUG found by round 2's failures: _SpanContext.exit compared identity against the context manager instead of the pushed _OpenSpan — every exc-info unwind no-op'd and leaked the span. _open now returns the _OpenSpan; exit pops by identity; ancestors still survive (GAP B).
• F4: _amortize_group_spans deleted (no fabricated per-context intervals, no _intervals writes). ONE amortization rule: group spans on the group ledger; per-context flat keys carry the context's own spans; per_item_amortized_ms is the single derived share via Ledger.batched_views (N5, now used).
• F13: per_item_end_to_end_ms = own prefill start -> own assembly end (note's definition); prior_ms reported separately; body amended.
• F9: dead timers deleted (_selective_second_pass, _build_schema_rows).
• F11: with-forms for prefill/rescore/reconciliation/dependency.
• N4: failed_attempts passes through the batched ScoreRowsResult; retry test asserts >= 1; second_pass_ms test is an equality against the dependency interval.
• Tests: 14 in test_w5b14_ledger.py. Gates: ruff clean, 733 passed (pre-existing api.py KeyError routed to W5-C: exact setcons solver, one count+set optimization, typed CalibrationBundle, internal telemetry, abstention contract #51).