[wshlavacek]

Pins reference-data-v3 so the nightly replays from it. It's the recording that #1325 needs: #1325 makes the RoadRunner SSA screen keep a slow model's reference on a recording, as #1293 did for rr_ode (#1290).

What was recorded

The record-mode run 40899940910, at e4255465 (the merge of #1325), recorded v3. Every suite passed and the night had no regressions. No entry was left out by the license gate.

The run's publish job was refused (HTTP 403) creating the release's tag at e4255465, on its first attempt and on a rerun. So reference-data-v3 was published by hand from that run's own assets, with the same title, notes and assets. Its tag is on main's head when it was published (05ee8039). This manifest was built by reference_archive.py manifest with the job's own arguments. That script is the same at e4255465 and on main. Each asset's sha256 here matches the digest GitHub reports for the published file. #1396 changes the publish job to tag the default branch's head, and explains the 403.

suite	v2 entries	v3 entries	engine
rr_ode	1,321	1,321	RoadRunner 2.9.2
bng_ode	592	592	BioNetGen 2.9.3
bng_ssa	110	110	BioNetGen 2.9.3
bng_nf	235	235	BioNetGen 2.9.3
amici_sens	96	96	AMICI 1.0.1
rr_ssa	218	223	RoadRunner 2.9.2

• rr_ssa. The recording timed out 14 models before their reference (v2: 19). No model joined that list. The five that left it are the five v2 added:

  case	v2 rec	v3 rec	v3 entry
  MODEL1102210001|ssa	TIMEOUT, no entry	PASS	yes
  MODEL2002030001|ssa	TIMEOUT, no entry	TIMEOUT	yes, saved past the cap
  BIOMD0000000800|ssa	TIMEOUT, no entry	DIFF (partial horizon)	yes
  BIOMD0000001013|ssa	TIMEOUT, no entry	DIFF (partial horizon)	yes
  MODEL2001160001|ssa	TIMEOUT, no entry	DIFF (partial horizon)	yes

  MODEL1102210001 is the one that passes on some nights. Under v2, any replay after a live night that passed it alerted. BIOMD0000000530, which replayed as archive_unreplayable under v2, has its whole ladder recorded and PASSes (vacuous_lowcount). Six rows ran past the 120 s cap and still saved their reference: BIOMD0000000558, BIOMD0000000767, MODEL1112100000, MODEL1909300003, MODEL2002030001 and MODEL8459127548. They stay TIMEOUT, as on a live night.

• rr_ode. Only the two giants, MODEL1504130000 and MODEL1603150001, timed out before their reference, as in v2. They don't alert: their baseline is TIMEOUT.

• bng_nf. One case that's good in the baseline wasn't PASS in the recording: "AD 3 State FREE Expanding nfs". It has an entry, recorded down the path it took. v2 had this case and "rm_tlbr".

Check

• reference_archive.py fetch resolved rr_ssa, bng_nf and bng_ssa from the published release with their sha256s checked. rr_ssa's archive holds MODEL1102210001 and BIOMD0000000530.
• This PR's nightly runs in replay mode against v3. That run is the check that the assets reproduce the live verdicts.