[W07] Restate reported numbers from the canonical run

[duckyquang]

Task metadata

Field	Value
Task ID	W07
Category	Writing
Priority	Medium
Complexity	Low
Est. days	2
Blocked by	E07, E08, E09, E10

Features / method

Replace transcribed estimates in the docs with the canonical result set.

Description

Once E07-E10 land, every number in the repo needs to be restated from the canonical run. Right now the README results table carries n=245 from an earlier cohort build, while the manifest in the repo has a 248-row test split, so the docs describe a cohort that no longer matches the code.

Also worth a pass for correctness while in there: the README and TECHNICAL.md both still describe the fairness and calibration analyses as part of the pipeline, and until E08/E09 produce artifacts those read as stronger than what exists.

Steps:

1. Update the README results table with canonical numbers plus CIs, and correct n.
2. Update the provenance note — drop the "transcribed from the E02 evaluation" caveat once it no longer applies, and state the device and command that produced the run.
3. Update TECHNICAL.md 8 with in-cohort solo baselines alongside the published ones, so the comparison is not misread as like-for-like.
4. Check off the resolved items in TECHNICAL.md 11 (pooling, pairing window, masking strategy) and delete any that E06/E07/D07 closed.
5. Add fairness and calibration results to the README, or state plainly that they are pending.
6. Confirm no synthetic harness-validation output from failure/demo.py has leaked into any results table. Those numbers look real and are not.

Acceptance criteria:

• Every reported number traceable to a run that was actually executed, with device and command recorded.
• No transcribed or estimated values presented as measured.
• Cohort n consistent between docs and the manifest in the repo.
• Synthetic demo output clearly labelled wherever it appears.