Publish the throughput pair an e2e delta is a ratio of

[iraj465]

Closes #426.

The defect

A journey e2e block publishes e2e_gain_pct and nothing else. A percentage is
a ratio; two ratios with different denominators cannot be added. So a consumer
holding only percentages is forced into a unit error.

GEAK already measures both numbers — ref_med and cand_med in
integrate_result.json, from the same A/B that produced the delta — and drops
them before publishing.

Evidence

/shared_nfs/hyperloom-claw, measured 2026-08-20, Hyperloom CI only.

Every KEEP block is missing the pair, and the naive sum is meaningless:

	n
KEEP journey e2e blocks on disk	36
...carrying a throughput pair	0
naive sum of their percentages	348.6 pp / 15 sessions

The denominator was measured and thrown away:

	n
accepted/stack overlays	26
...carrying ref_med and cand_med	25 (96.2%)
...whose pair reproduces its own delta	25 / 25
worst deviation	4.74e-03 pp

What the consumer can do once it has the pair — all 36 KEEP blocks replayed
through Hyperloom's real recorder and real collect_recorded_optimizations,
counting the arithmetic branch each credited step takes:

gain_method	before	after
baseline_chain — composable, points of one baseline	0	23
local_gain_projected — sums percentages	34	12
missing	1	1

The composable arithmetic is already fully implemented on the Hyperloom side.
It never fires because its input is not published. The 12 remaining are
live-path KEEPs from runs predating this change.

The change

One rule, three places:

Publish the two numbers the percentage is a ratio of, taken off the same A/B
that produced it.

file	change
interface/run_e2e.py — _journey_overlay_entry	read ref_med/cand_med from the same integrate_result.json the delta comes from, through the existing _ir_get so both flat and nested integrator shapes work
interface/run_e2e.py — _journey_return_entry	carry the pair off the accepted-kernel record
e2e_workflow/e2e_workflow.js	one e2eFrom(integ) helper on each accepted head/kernel push, so the live path carries it too

Field names are base_tput / new_tput — the names the consumer already
reads. One name end to end, no rename in the middle.

A run with no measured pair emits null. It never fabricates a denominator.

Tests

interface/test_run_e2e_dispatch.py::TestE2EDenominatorIsPublished, 4 tests:

test	asserts
..._publishes_the_pair_and_it_reproduces_the_delta	pair present, and new/base reproduces e2e_gain_pct
..._reads_the_pair_from_the_nested_shape_too	nested integrator output is not silently dropped
..._return_path_carries_the_pair...	live path carries it
..._a_missing_pair_is_null_not_fabricated	no invented denominator

	result
new tests without this change	4 / 4 fail
new tests with this change	4 / 4 pass
test_run_e2e_dispatch.py + test_run_e2e_recovery.py	205 passed

Scope and limits

• The JS change is reviewed by reading only. There is no JS runtime on the
  analysis host, so e2e_workflow.js was not executed. The edit is 11
  mechanical call-site substitutions plus one helper; the diff is deliberately
  uniform so it reads at a glance. history.ledger sites were deliberately
  left alone.
• This does not make per-session totals correct on the replayed data. 10
  alias twins across 9 of the 15 sessions still double-count. That is the
  producer bug kernel_journey writes one acceptance twice when the overlay tag and the kernel symbol differ #407/fix(journey): fold a return-named acceptance into its overlay entry #408 fixed (merged 2026-08-19); every replayed session
  predates it. Not reintroduced or worsened here.
• One variant shape is left alone. Llama-3.1-8B-Instruct/20260816T114410Z
  → cand_pa_ll4mi_c0_hip publishes e2e_delta_pct_all_repeats (1.330) and
  e2e_delta_pct_equal_count (4.298) and no plain delta, so e2e_gain_pct is
  already None there today. Choosing a canonical variant is not something the
  data settles. Noted in kernel_journey e2e blocks publish a percentage with no denominator, so gains cannot be composed #426 as its own follow-up.
• Visible, not yet sized. Hyperloom must also stop discarding these records
  (fix(breakdown): put GEAK's kernel journey on the canonical streams Hyperloom#1252). Both are needed before the dashboard's
  Kernel Agent Gain (GEAK) column is both non-zero and correct.