Zero-E branch derives its own temperature block

[dfalster]

set_leaf_states_rates_from_psi_stem's zero-transpiration branch wrote
Tleaf_ but re-derived neither the Farquhar block nor the leaf-to-air
deficit at it -- while ci_ is read off gamma_ and assim_colimited_ off
the whole block. The transpiring branch re-derives per candidate by
design, so this one inherited the last psi the optimiser probed, making
the objective evaluation-order dependent. Both are derived now, before
ci_ is read.

Moves results on the PM path, the ProfitMax argmax among them (0.5.0 ->
0.5.1, NEWS has the table). Closes #110, #112.

[dfalster]

What was wrong

set_leaf_states_rates_from_psi_stem's psi_upstream >= psi_stem branch is the no-flow / reversed-gradient case. It wrote ci_, transpiration_, stom_cond_CO2_ and (since #106) Tleaf_ — but re-derived neither the Farquhar block nor the leaf-to-air deficit at that Tleaf_. Three things then read a temperature the branch had not seated:

• ci_ = gamma_ * umol_per_mol_to_Pa_ — the compensation point is a member of the temperature block;
• assim_colimited_ = assim_colimited(ci_) at the bottom of the function — the whole block;
• vpd_leaf_, which is a reported field and is what g1_eff() divides by (The shut-down branch leaves vpd_leaf_ stale #110).

The temperature it read was not the air baseline. That is the part that made #112 separate from #105 rather than a repeat of it: the transpiring branch calls update_temperature_dependent_params per candidate by design (it bypasses the photo_temp cache — Tleaf varies per operating point), so the zero-E branch inherited whatever ψ the optimiser last probed. The objective's value at a fixed ψ therefore depended on the order candidates were visited.

Reachability, since the issues asked for it first

Four routes, established by reading the callers and confirmed in the probe:

1. prepare_profitmax's scan, at i == 0, on every run. The scan starts at psi_soil and calls set_leaf_states_rates_from_psi_stem(psi_soil, psi_soil) — equal potentials, so this branch. Its A is recorded and is a legitimate profit candidate (that is hazard 11 / Leaf-to-air VPD, and Sperry's ProfitMax #93's grid argmax), so this reaches an argmax, not just a report.
2. profit_at_collar_psi, which clamps a caller's target into [bound_a, bound_b] where bound_a is root_zero_E — psi_stem == bound_a exactly there. So the branch is the final reported state on the prescribed-collar path (TF24f's acclimation route).
3. maximise_profit_over_collar's two off-normal exits — the golden-section fallback and the two-pin SolverRefused case — both evaluate profit_psi_stem_TF at lo/bound_a.
4. The standalone optimise_psi_stem_* optimisers at their wet endpoint.

find_root_collar_psi's normal path does not reach it: dprofit_at_collar_psi has its own psi >= psi_stem exit, and the collar it returns is stepped inside the wet bound. The assim_max_ < 0 branch #112 originally paired with this one is gone — #93 deleted it, for a better reason.

The measurement

Probe built against master and against this branch from separate trees (a git worktree --detach, not a stash — git stash cannot revert a commit, and a half-reverted build reads as a real finding).

The state the branch reports. Tair 30 °C, PPFD 900, ψ_soil 1 MPa, D_air 1.5 kPa. The branch's own leaf temperature is 39.325 °C against Tair 30, so there is a large gap to get wrong:

reported at the zero-E point	cold object (Tair baseline)	after one transpiring candidate	this branch
assim_colimited_	−1.9888	−2.5355	−3.0978
R_d_	1.9888	2.5355	3.0978
vpd_leaf_ (kPa)	1.5000	2.7368	4.3714
ci_ (Pa)	5.5700	6.9492	8.7173

Two columns, two different histories, same call. Respiration was 36% low and the deficit 66% low at the worse of them, both in the direction that makes a shut leaf look cheaper than it is — the same direction #105 found, for the same physical reason (no latent cooling, so this is the hottest the leaf gets).

The ProfitMax argmax moves, in both directions. Thermal cost on, PPFD 900, D_air 1.5:

Tair	ψ_soil	argmax before	argmax after	profit before	profit after
40	0.5	0.5000 (shut)	1.6052	−1.2825	−1.2309
40	1.0	1.0000 (shut)	1.7595	−1.3032	−1.2983
40	2.0	2.0000 (shut)	2.0000 (shut)	−1.3351	−1.3657
50	0.5	0.5004 (open)	0.5000 (shut)	−1.8720	−1.8720
55	0.5	0.5004 (open)	0.5000 (shut)	−1.7755	−1.7754

At Tair 40 the closure candidate's carbon was too generous, so closure won when it should not have; correcting it hands the interior point the argmax and a higher profit. At Tair 50–55 it goes the other way — the objective is nearly flat there and the answer becomes exactly shut. Neither direction is a rounding difference.

The prescribed-collar path, clamped onto the wet bound after a real solve had warmed the block:

Tair	profit before	profit after	vpd_leaf_ before	after
25	−2.4329	−2.5415	3.4861	3.7486
40	−3.84281	−3.84282	6.0536	6.0538
50	—	—	(unchanged: takes the shade-death exit, already fixed by #111)

find_root_collar_psi is unmoved at every PM driver tried — the checksum over four soils at Tair 40 is bit-identical across the two arms. Stated as what it is: a statement about the drivers tried, not a proof, since routes 3 and 4 above exist.

Cost (hazard 5)

This adds two transcendental blocks to a branch of a function the collar solve runs ~10³ times per solve, so it was measured rather than assumed cold. Both binaries built from their own trees, interleaved ×4 at load ~5.4, min-of-400:

arm	before	after
zero-E branch + one transpiring candidate, PM on (µs/pair)	4.972 / 4.990 / 5.031 / 5.178	5.018 / 5.012 / 5.205 / 5.218	+0.4% … +3.5%, median +0.9%
find_root_collar_psi, PM on (µs)	249.2 / 251.3 / 258.0 / 255.5	247.3 / 243.2 / 256.8 / 258.3	no signal
optimise_psi_stem_ProfitMax, PM on (µs)	2647 / 2645 / 2680 / 2710	2620 / 2627 / 2707 / 2717	no signal
find_root_collar_psi, PM off (µs)	84.66 / 84.72 / 88.14 / 88.02	84.65 / 88.10 / 88.10 / 88.81	no signal

The round-to-round scatter on the last three arms is wider than any difference between the arms, and the per-pair figure is dominated by the transpiring candidate that shares each iteration. Gate-off is unchanged by construction as well as by measurement: the recompute is inside if (use_energy_balance_), and set_leaf_vpd returns atm_vpd_ exactly there.

Evidence and its limits

• Both golden files bit-identical, and that is not evidence here — both grids run gate-off, where there is no second temperature to disagree with. Same for leaf_behaviour_fingerprint(), which is golden-derived. This change signals only through NEWS and the version bump (0.5.0 → 0.6.0).
• 670 C++ checks, 0 failures; make bench builds; the R testthat suite is green including gradient_golden.tsv bit-exact.
• The new test's sharp assertion is order-independence rather than a value: the same (psi_stem, psi_upstream) on a cold object and after a transpiring candidate must agree bit-for-bit in Tleaf_, assim_colimited_, R_d_, vpd_leaf_ and ci_. A value assertion would not have caught the defect this change is about, since the value was plausible at either history.
• One assertion had to be weakened and it is worth recording why. assim_colimited_ == -R_d_ is exact at the PM shut-down reports Tleaf but R_d at air temperature #105 exits, which assign it; here it comes through assim_colimited(gamma*), where gross assimilation is analytically zero but the co-limiting expression rounds — measured 4.44e-16, one ULP. The bit-exact statement that survives is against the model's own function at the model's own ci_, which is what pins the temperature.

What this does not do

optimise_psi_stem_TF and optimise_psi_stem_Sperry still have hazard 11 (#94), and the surface question #114 raises is untouched.